✨｜sdxl

[uncut steeple]

[jagged compass]

Hello people, I have a question, I want to use SDXL, but I don't want to run it on my hardware, besides clipdrop, what are my options?

[tepid maple]

vastai or runpod have a look at them

[uncut fiber]

isnt there some solution unlimited pictures 7USD/month? or is it clipdrop?

[tepid maple]

clipdrop actualy good pricing

1500 image a day is really good

imo

makes 45k month

it is good

[peak dove]

ComfyUI 1221 too old - where do I get an update from at all?! 🙂

[tepid maple]

go to github tor

[peak dove]

Github - OK, thanks

Cannot tell which version at GitHub?

[heady vale]

use the update bat file in the comfy folder

[clever verge]

There’s only main version, you could just use git pull in the comfyui directory

Or run the bat as above, guess it does the same thing plus updating the python modules.

[peak dove]

I am using Manager to update - even after an update this way - still says 1221 is an old version?!

OK, .bat file

So ComfyUI Manager does not actually Update ComfyUI?! The mind boggles!!!

Which .bat file, to be specific?

OK, found it 🙂

Found 3 items not on PATH ...

Prompt Queue - is it now hard-wired to the right side of the screen?

[peak dove]

Moobs!!! And what beauties they are!!! 😄

[peak dove]

How do I make PROMPT QUEUE in ComfyUI free-standing again?!

[uncut fiber]

1330 is latest version? @peak dove update it as Pure-fire suggested.

[peak dove]

Yes, I discovered the .bat update files - especially confusing as ComfyUI Manager suggests that the way to Update is via Manager - yet I've had to do it manually ... ! I've learnt something at least 🙂

Some results this a.m. on newly updated CUI

[wet nacelle]

Not sure yet.

[eternal fog]

Coil whine shouldn't be an issue. It's just components humming at a frequency you can hear.

GPU manufacturers seem to have gotten lazy in recent years dampening them.

[wet nacelle]

https://tenor.com/view/bebezinho-little-baby-crying-sad-tearing-up-gif-15419141

"Clark Kent It's me the Bat Man."

[eternal fog]

I think the "fax machine" noise is just down to how it loads the GPU. Mine does the same.

I think each step is a short high load on the GPU. So you get the buzz-buzz-buzz sound for each step.

[safe stump]

Hi

[wet nacelle]

Hello

[safe stump]

Hey can i use stable d.. Model in my app if i don't brake the usage agreement

[shy kelp]

mack do you use comfy ui? can i have your workflow

[wet nacelle]

Drag and drop any of my images into the UI itself.

[shy kelp]

aah i see, thanks mack

[wet nacelle]

yee

[shy kelp]

do you know how to remove previous load photo list?

[wet nacelle]

nope

@cold cargo Hey how are we on tasteful nudity? My nudity gen was by chance and I really just like the way it ended up.

\

[shy kelp]

how do you combined those photos

[wet nacelle]

With weights

[shy kelp]

using clip, any tricks?

[wet nacelle]

yeah just use clip

[wet nacelle]

I don't combine. Combineing to me is still re tardedly broken. I can't figure that out yet.

I use clip just for art styles right now.

[shy kelp]

do you fill all the box with image?

[uncut fiber]

not sure i will ever learn this...

[shy kelp]

haha

[shy kelp]

sooo, you are using auto11?

[uncut fiber]

No comfy, just very basic workflows 🙂

[elfin gull]

/

[shy kelp]

haha cute

[uncut fiber]

dont know anything about image input, so combine it, its behind my horizon 😄

[indigo carbon]

which is best?

[upbeat summit]

the first one is very different of course. but coherence is best in no 3

[indigo carbon]

they all used the same gen params, only difference is the SDXL finetune checkpoint

[upbeat summit]

well than in my case I would further explore the checkpoint of no 3

[wet nacelle]

@shy kelpTry to undesrand it like this. I use three images that all have the same art medium becasue I want realism. I do not mix mediums (for right now) because I have found worse results when trying to even out the weights.

I use three images becasue I found that I get the style I want more often by having that third image. I include the fourth image because I might try to see what I can do with it one day in regards to making it look good.

If you want to blend images without a prompt then you will need to use the "strength" value. When blendeing you want to mentally understand what you want to see the most of. I don't often go above the values of 2 because it doesn't mix well on the putout from what I have tried.

Moving onto prompting with the images. When using the posative prompt you want to describe either want you want to see that involves the images you have chosen or any other subject at all.

When I first started using it i tried to use the prompt to get what I wanted to see from the images. I quickly learned that doing that wasn't always the best idea. I now only use it for making a style. I grab images that are photos of faces and sometimes locations because I want hyper realism.

And if you want that fourth image then click the nodes that have the purple covering it and press ctrl + b. This is bypass those nodes. When you have them back on you then wire the "unCLIPConditiong: "CONDITIONG" output to the "KSampler" "positive" sinput. You'll also want to rewire the last node of the unCLIPconditioning so that the "CONDITIONING" is wired to the "positive" input on the "KSampler"

Pay attention to where those wires are going so that you don't get confused when tinkering with it all.

Be sure to save a lot when you don't really know what you are doing.

[uncut fiber]

it need undo/redo 🙂

[shy kelp]

wow mack, nice thank you so much for the tips!

[shy kelp]

yea agree, as 3d artist it is like blender, i hope i can layout the keys as blender,

[indigo carbon]

which one this time?

[shy kelp]

i like on the right one

[uncut fiber]

snowboarders

[wet nacelle]

Yeah no problem.

[noble shoal]

So, i think i have found a solution to my Reroute Switch problem. As it turn out, that you can't mute a reroute node to break the process chain and you need to make a switch out of 2 reroute nodes to connect them to make a switch. i have figured out another way by making "passthrough" notes. Maybe it was obvious, and someone figured that already out, in this case feel free to call me dumb 😄 . Those "Passthrough" Nodes can just be muted with ctrl + m. And they break the chain and stop anything after them from processing.

[upbeat summit]

that is not dumb - it's great that it works for you. I've been using the "jumper" technique for some time. some nodes can be muted, others can be bypassed

[noble shoal]

Yeah, i basically get rid of my jumpers now

[shy kelp]

crazy good

[indigo carbon]

I know I sound like the SDXL beta bot, but A or B?

[peak dove]

b its grungier

[indigo carbon]

A is CCXL and B is copax_colourful

[peak dove]

A

[peak dove]

Cool!!!

[indigo carbon]

idk which I like better, both are very detailed and unique

both are SDXL though, so my final conclusion is whichever is more coherent will be better imo

[uncut fiber]

is dynavision any good?

[peak dove]

ComfyUI - I used to be able to move the Queue Prompt to any position I liked - now it seems stuck to right-screen?!

[indigo carbon]

I kinda like copax_colourfulv2 rn. for me it's a tie between CCXL and copax_colourfulv2

[uncut fiber]

this is batch of dynovision, good it all had moreles well made standing on snowboard.

[peak dove]

Totally cool!

[indigo carbon]

try the copax colorful model, I tried everything, the copax model has the highest average aesthetic score

[uncut fiber]

o.k. thank you @indigo carbon will try!

[peak dove]

Mr Trump drives himself to Fulton County, Georgia ... comfyUI

[uncut fiber]

@indigo carbon there is v 4.2 some numbers version and then OK going with colorful V2

[indigo carbon]

one thing I notice, copax_colorful seems to be slightly grainier than CCXL sometimes

possibly requires higher denoising strength with refiner

[noble shoal]

You all with your highres images. And i am here with my newly trained 32x32 pixelart lora 😅

[indigo carbon]

I don't really get pixelart tbh, can't you just make normal images then downscale?

[azure oxide]

firefox says this is a 747x598 image tho

[indigo carbon]

can't they just take a normal AI image, downscale it massively, then upscale using math(no upscaler, just resize)?

[ruby sable]

Pixel art has to do with very deliberate placement and colors of pixels, chosen by an artist for aesthetics.

Basically it's very stylized. So downscaling won't get you there.

[indigo carbon]

nah, worst case scenario just make a LoRA for that

[ruby sable]

Yeah that's a way

[noble shoal]

@indigo carbon Here is what you get with your downscale upscale techique 🤷

[indigo carbon]

that's how pixel art usually looks like..

it seems to have worked

[noble shoal]

I appreciate you, Tdg8uU, for giving your respect for other artstyles. Thank you for your valuable contribution in this conversation.

[soft zealot]

I agree a proper generic "switch" would be nice.

I m not a big fan of muting simply as I always seem to ignore the greyed outedness of them lol.

Ah well stick with my "Enablers"

[upbeat summit]

Does anybody know some tokens how to prevent jewelry chains in realism photography? They keep popping up and are rarely coherent.

I tried chain, chains, jewelry chain, metal in the negative or wearing no jewelry in the positive

[uncut fiber]

@upbeat summit how to put on jewelry or jewelry chain, more attention in negative prompt?

[upbeat summit]

(neck chain, jewelry, jewellry, jewelry chain, jewellry chain, chain, chains, snake chain:1.3)

trying this in the negative right now. it has improved it and I get way less chains now. before it was like 2 -4 or more - now it's mostly down to 1

[uncut fiber]

or (jewelry:0.1) in positive

[upbeat summit]

good idea

[uncut fiber]

i am using 1.5 and it is still somehow o.k. don't broke prompt very

Tried to make space suit helmet cubic shaped, or not wear it and it still does same suit with same helmet

[upbeat summit]

adding (jewelry chain:0.1) in the positive adds massive amounts of chains to the image 😄

the mysteries of the latent space

[uncut fiber]

ou weird, probably [jewelry:0.1]

but weird weird

[noble shoal]

(jewelry:1.5) in the negative? Maybe

[upbeat summit]

in ComfyUI you use (brackets)

[uncut fiber]

o.k. not squared brackets or how they are called? Going to try on my own 🙂

[upbeat summit]

tried it - it does reduce the amount of chains though but you can't get rid of them. might be the overall prompt build though

[uncut fiber]

Waterproof - not

@indigo carbon using yellow. Moving on some other theme 🙂

[indigo carbon]

yo?

[soft zealot]

so a short while back (2nd june to be precise) I got a load of data from the test results at https://vladmandic.github.io/sd-extension-system-info/pages/benchmark.html (nothing to do with me) to draw up a Relative Performance Ranking of various GPUs.

Data for the 4060ti was added on 25 July.

Again I want to stress I'm more interested in Relative performance rather than whether or not the absolute it/s are correct for each GPU (and again all the names come from the site referenced, not my data)

Should also be noted that these would have all been tested/created using SD1.5 or 2.1 models but again, its a relative ranking.

So for no particulr reason I've now also splitthese into Three Tiers with each Tier subdivided into a further 3 tiers

[uncut fiber]

that you are using orange space suit 😄

[eternal fog]

This doesn't work. Putting jewelry into the positive at any strength will add jewelry more than it not being there.

[visual glade]

benchmarks on a1111 are invalid because the unet code has autocast on

[indigo carbon]

oh hey comfy

[soft zealot]

as I said its all relative rather than absolute

in other words I wdont necessarily disagree however if every test is wrong to the same degree the relative positions will remain the same

[indigo carbon]

@visual glade can you push the compile AIT code to the custom node? we can't seem to figure out batch size with AIT

[visual glade]

which error do you get when you try to compile with different batch size?

I set batch size as a IntVar(values=(1, 16), name="batch_size") and it works

[indigo carbon]

msbuild errors, hold up- I'll find the exact errors

[soft zealot]

consitency over accuracy 🙂

[visual glade]

but there's a few things wrong like the H100 should be on top

and the different A100 should be pretty much the same

[soft zealot]

I just crunched the numbers. It's not my data and I have no bias 🙂

[uncut fiber]

how comfui is working, is it difficult to add undo redo?

[soft zealot]

@visual glade would you consider building a standardised benchmarking node in COmfyUI that autofed backinto a DB that was publically accesible

[crisp owl]

https://github.com/bmad4ever/ComfyUI-Bmad-DirtyUndoRedo

[uncut fiber]

@crisp owl thank you!

[crisp owl]

It ain't perfect, but works when you need it

[steady grove]

neato

[upbeat summit]

I don't know if it's fixed now, but in an earlier version that I've tried if you use Ctrl+Z + Ctrl+Y (undo/redo) and are in prompt text boxes you also reverting to an older version of your workflow. They were not excluded from it. And I use undo/redo a lot when doing prompt engineering.

But I guess text input boxes could be technically excluded from it

[crisp owl]

I never looked into the code behind it, but from what I understand it basically keeps a cache of each change you make, and that's what it's bringing up each undo.
Something along those lines.

[upbeat summit]

yes this is how it works

but CTRL+Z must be excluded from text input boxes, otherwise you are reloading an older version of your workflow every time you want to just undo a change in your prompt

otherwise it works - but this was a deal breaker for me. if it's not fixed yet, I will take a look at the code, because having an Undo function would be great of course

[crisp owl]

Yeah a simple undo button would be phenomenal, but good things come with time haha

[indigo carbon]

currently trying this. is set AIT_USE_CMAKE_COMPILATION=1 required for batch sizes?

[visual glade]

oh yeah I set that to 1

[autumn forum]

True, I’m really looking forward to the future of comfyui. Hopefully it gets a little more. Comfy😛

[indigo carbon]

thanks, will update if compiling succeeds or fails.

[timid garnet]

RIP the rabbit

[visual glade]

for me compiling anything just failed completely without that on windows

[crisp owl]

Well, it's already worlds more customizable than other platforms, and for that, I'm hooked

[indigo carbon]

heh, for me it worked without it. but it appears to be required for BS

yeah it's compiling, don't see any issues yet

[upbeat summit]

@indigo carbon has anything new been found out how to accelerate LORAs in combination with AIT?

[indigo carbon]

it works, it goes- model->LoRA->AIT->WORKFLOW

@visual glade how long did it take you for a BS 1 16 module? when I compiled for BS 1 a few days ago it took over an hour

[visual glade]

comfyui patches the model weights with the loras so it should work with AIT

didn't take too long for me, maybe 10-20 min

[indigo carbon]

jesus, did you do just the UNET or something? when compiling the entire model it takes an hour or 2 lol

[visual glade]

yeah I only do the UNET

[indigo carbon]

that explains it

well, it's still going, it's now doing text encoder2

wait, is there a --flag for just the UNET?

[upbeat summit]

it does work since the beginning but it takes 4 and up to 10 seconds every time before it starts generating an image if you have a LORA in the chain. model->LoRA->AIT->WORKFLOW

at least on my system but others have experienced the same. Adding this much time (it's like frozen) for every gen is of course not optimal.

If I bypass the Load LoRA node the speed is back to normal and the delay is gone

the delay is also gone when using --gpu-only on start up, but this tends to get really unstable after a couple of minutes on my system (freezes and slowdowns)

[peak dove]

My ComfyUI is going crazy. Started off at a good s/it rate ... then really slowed; then added in all by itself an Upscale (which node was completely disconnected to begin with!!!) Ghosts?

[indigo carbon]

oop it failed

@visual glade any idea?

[upbeat summit]

And I don't really know if this is an issue that I report to the ComfyUI repo or the AITemplate repo heh

[indigo carbon]

it went smoothly at first, when it finished the 2 text encoders it started UNET and spat out this error. this error doesn't occur when doing bs1

[thorny frost]

has somebody tested anty of the Kappa_Neuro Loras at Civitai?

They doesnt seem to make any effect at all!

i loaded a lot of styles and none seems to work

is it some kind of scam?

you can add or delete the lora of the prompt and it will no affect to the final result at all.

[uncut fiber]

aquarium fish for dish 🙂

[crisp owl]

Which one? Is it for SDXL?

[steel mist]

Not sure what the scam would be. I mean, the images definitely look AI generated

this hand was not painted by a rococo master heh

[thorny frost]

he has versions for 2.1 and XL. I am trying XL

[thorny frost]

the images are AI generate, but you obtain the same image with the lora and without the lora. It looks rococo because the style is already in the base model

they all definitely looks like empty loras. Most absurd thing i have seen.

[crisp owl]

Oh kappa neuro is an account name, I was looking for a lora named that lol

[thorny frost]

yes, he has a lot of styles. Pretty amazing if they were real

[uncut fiber]

two submarines for cats

[thorny frost]

i doubt about the watertightness of those designs though. :p

[thorny frost]

https://civitai.com/user/Kappa_Neuro/models

[crisp owl]

Oh well he's got a few

[steel mist]

Yeah I mean the volume would make me wonder what they're training all that on?

[uncut fiber]

i have no doubts 😄 i am sure it wont work 😄

[steel mist]

Where does the default config hide for comfy?

I keep getting tripped up with the default being 512x512

[thorny frost]

I think that is a node thing

[hardy cipher]

yeah, shouldn't be hard to find

[thorny frost]

On nothing. So i downloaded 30 gigas de crap. Deleting...

Most silly scammer ever

[steel mist]

lol you're not kidding. it's not the same image it is pixel exact

I had downloaded some of these but hadn't played with them

[thorny frost]

yes those loras are like ghosts

[autumn forum]

Lmao that’s crazy, have you tried turning it to wumbo?

https://tenor.com/view/wumbo-spongebob-gif-10902129

[steel mist]

I was looking forward to the craig mullins one

Loved him back from the marathon box art days 😂

[thorny frost]

you can for instance change the loras weight to any number with zero effect

[thorny frost]

lol

[hardy cipher]

xyrrus: nodes.py lines 862/863

in the main comfy folder

[proper depot]

Any advice on a sampler to use if I want to render out some final, print-worthy pieces? 🤔

[thorny frost]

So that is the secret of Bill, he can count up to 12, two more than us.

[steady grove]

have you used their tokens?

if they're trained with good regularization, they won't come on without tokens

[hardy cipher]

yes, he is 20 percent better at everything, including counting

[thorny frost]

That is the thing. You can use the tokens without downloading the lora of civitai and will get the same result

[steel mist]

I used one of the prompts directly from a sample

identical results with lora/clip 0 and 1

Also given the sheer volume of these

it raises a question

[thorny frost]

Is the guy tring to get some record uploading trash?

[steady grove]

yeah i went and tried it. first test got unstunning results. looked at the prompts and the prompts are doing all the lifting

might just be enthusiastic and thinking his loras are doing it all

[thorny frost]

or maybe he dont know what is he doing.

[steel mist]

strength 0 vs strength 1, prompt copy paste from a civit example

0 difference

[steady grove]

trains for 5 steps. fires off a very styled prompt. calls it done

[steel mist]

a decent finetune takes what, 10 hours on a 4090?

[hardy cipher]

there has to be a more efficient method

well, maybe not known

[steady grove]

for a style lora? naw. if you're blowing out your memory then yeah over 10. smaller batches so that the vram doesn't over flow into shared helps a lot. Smaller batches does better with gradient accumulation which takes time, but takes less time than blowing into shared memory pools

[hardy cipher]

but someone will come up wiht some kind of magic

[steel mist]

Just talking practically about how many this dude could theoretically crank out

[thorny frost]

I have just seen some Magic in LLM with llama.cpp

[hardy cipher]

could be using cloud

[steady grove]

2 hours for a style lora first version i'd say. If its popular then do it better. if i were trying to mass produce that's what i'd do.

this guy though. kind of sus. neat find.

loras don't seem to be doing jack at all. i've tested 3 now

[thorny frost]

he has fun uploading these things apparently

[steel mist]

Yeah if they seemed to work I'd think "sure he's training on cloud and just loves art"

but since they don't work, and there are a ton of them

🤔

[steady grove]

🤔

[steel mist]

Given the amount of crap on civit I guess I should not be surprised

"mind blowing realism lora!" and the example pictures are plastic skin with 12 fingers per hand

[steady grove]

https://ko-fi.com/kappa_neuro heres the grift

has actually caught some

[hardy cipher]

well even outside of training time entirely

how do you find all those pictures and put collections together like that?

[steady grove]

wonder if each file has the same hash

[thorny frost]

gotcha

[steel mist]

the hashes on civit are at least different based off a quick spot check

[steady grove]

no its different each time

meta data can do that. i bet it's the same lora each time though

civit has remote upload too iirc

[steel mist]

well if you report it to civit let us know if we can +1

[thorny frost]

how can you see the hash od a file in wiondows?

[steady grove]

yeah with tools. i used civit though

[steady grove]

i think civit doesn't really facilitate reports too well since the whole ai trolls would use them

just this

[peak dove]

Gill Bates?! 🙂

[steady grove]

i threw up "needs moderator review" and explained that i think the styles are automated prompts and the loras are blank templates

i honestly think they dn't read them though. How many of these do you think they get?

[thorny frost]

soemthing like that?

[peak dove]

We Are Borg - or We Are Tylda Swynton!

[hardy cipher]

they're all the same size, lol

[hardy cipher]

wait, one's 435.6

[thorny frost]

they are different sizes

but same result

[hardy cipher]

well clicked like 8 of them

435.35, 435.36

[steel mist]

~can you guess which ones are from this guy?~ nm there are ones from other authors with the same size

[hardy cipher]

to be fair, I didn't look at all o fthem

[steady grove]

I had explained the prompting were doing all the style work. I think that's a key since it appears in the examples at a glance that it works

[steady grove]

there's hundreds i estimate

[hardy cipher]

so is he cheating or just exploiting what's allowed?

[steady grove]

looks like bot. all his posts have hundreds of likes.

[hardy cipher]

ahh

you know, that's the part that confused me

[steady grove]

hundreds of mundane style loras all being hyped so hard? i unno

[thorny frost]

"all kappa_neuro loras seems to be empty. It is the prompt who does all the style work. The Loras does not affect at all"

[steady grove]

coo

[hardy cipher]

and saw some of them on a "featured" list

[steady grove]

i also said i reported based on this discuussion on this server

[thorny frost]

reported

[hardy cipher]

lol, well whoever is doing it created a very prolific system

[steady grove]

https://tenor.com/view/band-the-muppets-rock-out-rocking-out-i-dont-like-spam-gif-5375842

[thorny frost]

the guy has been uploading nothingness for months

[hardy cipher]

well I wish they didn't suck

some of those look pretty cool tbh

[thorny frost]

will try with one 1.5 to see what happens

[steady grove]

i think he's using the one button prompt script. all his examples have automatic1111 metadata

it's pretty good for doing random artists and styles

i wish it was stand alone actually

[hardy cipher]

this is why people can't have nice things

[noble shoal]

So he has just wasted 47.5 tb of civitai's traffic since today.

[thorny frost]

i have used one button for months and having lots of fun. Didnt imagine it could be used to do evil things. LOL

[steady grove]

i've been using it pre work. i'll go in an theres a mode to generate prompts without using them. do that, pull the list into a txt file. take it into comfy.

[steady grove]

scammers lovvvve to automate

[hardy cipher]

https://tenor.com/view/think-smart-gif-22411928

[thorny frost]

well, he at least has shown us how amazing the base model is

[steady grove]

leveraging people's blind spots for donations

[hardy cipher]

grimy

[thorny frost]

civitai should curate content a bit better. And clean some hentai crap too

[steady grove]

grew up in a tourist city. there'd be guys who walk the block "scuse me i don't have money for the bus, could i borrow $2, $3?" and then bam, move on. they do this 300 times and 10 people help em out because its a good story and a person needing bus fair is more helpable. blind spot beggers.

most street beggers weren't even homeless people. lease cars and condos and shit. a good day can see $1000 cause tourists all think they can make a difference

[hardy cipher]

man, just make a "beware, degenerate content" section or soemthing

[steel mist]

There’d be nothing left 🙃

[steady grove]

real homeless people look too hard up to be allowed downtown. police take em on a trip to the shelter on the other side of the hill

[hardy cipher]

there was a guy that'd work mlb games near me. follow people along, begging begging begging like he was a crack addict. just pathetic. then when he'd get done he'd go back to normal

get in hius car and drive home

[steady grove]

there's a ton of good will in the world and the people who are most generous have a lot of blind spots

[steady grove]

seriously it's just an act a lot of the time. 9-5

[hardy cipher]

a lot of times people don't pay them because they feel bad, it's more to get them away

if you're following some family with little kids or something begging

not like the parents want to deal with that

but yeah, with the lora thing

people just don't know any better

[peak dove]

Some recent ComfyUI SDXL

[hardy cipher]

there's a lot going on. can't really expect people to know the intricacies of how traing and all that works. I've been using SD quite alot for about 9 months, and I stiill don't know how a lot of it works

[steady grove]

another city near me, it's less of a tourist city. guys will drive around picking up homeless people and then take them to highway exits and intersections to beg or wash windows. these are like wayyy out in the woods with no facilities near by. keep these guys hydrated and fed then take half their "earnings" . they're basically homeless people pimps.

it won't be long till ai waifu pimps show up. these sort of advantage seeking people love to organise hierarchies and cults

[hardy cipher]

lol, that's messed up

[steady grove]

its not the real messed up part. to shine light on just how much of a human traffiking operation it is, you'll notice that they never have women at the intersections begging. Women can generate cash for these operators in other ways.

[thorny frost]

The guy keeps uploading no-loras at one every hour aproximately

and in alphabetic order

[steady grove]

anyways. long story short, don't underestimate people's capacity to abuse good will

[thorny frost]

we are in Jo- now

[hardy cipher]

not sure it's a thing there, and not disparaging anyone, or ethnicity or whatever. but do you have any of the "roma" folk around there? because there aer a few that show up around here. and they all have the exact same approach

[steady grove]

the first company that trains an AI classifier for data intake that can be used at an enterprise level has a huge position

[hardy cipher]

signs with identicdal hand writing

[steady grove]

something that can determine scams like this at intake

[hardy cipher]

show up, I mean like one lady will show up to big events with her kids and a sign with very clean, perfect hand writing

but then, saw another lady with the exact same hand writing on her sign

and then a post online in another city, same hand writing

[steady grove]

if i were on the marketing team, i'd call it "hopper" and the logo would be an obey style stencil of dennis hopper

[hardy cipher]

makes me wonder if there's a guy. the sign guy

and he just makes signs

[thorny frost]

well the 1.5 lora seems to work

[steady grove]

well, i mean, gypsies kind of come from romania no? there's a lot of stereotypes about gypsies for many reasons.

I kind of love their culture. They're kind of grifters but they seem more like robin hood an his merry men about it.

its about familia

[hardy cipher]

yeah, I just navigate some things with caution. not sure who might get offended by what

I have no hate on them. I'm just impressed that I've seen the same woman with her kids at least 100 miles apart

kind of low to exploit her kids like that

[steady grove]

i used to always say gypped thinking it was just normal. had no idea it was a gypsie slur for a long time.

[hardy cipher]

but might not be her call

well they really have it figured out

you want to look down trodden, but not low class and scummy

just like you lost your job a couple months earlier and you're running out of options or something

[steady grove]

a good sign is powerful! the art of sign making is something homeless culture runs with. i'm not surprised that those trying to maximize the begging career are using mass produced signs

the key is a sign that makes people actually read it or look at all

tougher to do than it sounds

[hardy cipher]

well it's all the traveler folk with these signs though

maybe they have a sign syndicate we're not privy to

[steady grove]

they probably printed off 100 and got a pile of em

"hand written"

[thorny frost]

the same lora in the xl version doesnt make any difference though

maybe the guy simply doesnt know how to train XL

[steel mist]

And doesn’t know how to do basic testing?

[hardy cipher]

well I've just seen the same ones in dallas and houston. and then I saw something online inn like st. louis or something, same signs. LOL

kind of blew my mind for a minute

[steady grove]

naw i think he did train loras on styles the models already knew before. easier an efficient. probably used a collab to automate it too.

collab stopped working. now he's got a new script deployed.

[thorny frost]

yeah, he is too silly or too clever

[steady grove]

it is a syndicate. a homeless one. these "homeless pimps" don't operate in just one city. they take their gems to the hot spots all around the country. a mother with kids for instance

or just a bunch of kids in general. probably not even related. human traffiking is so common unfortunately

[steady grove]

the city i grew up in, i used to volunteer a lot of homeless outreach. going an talking to them basically. i'd have pamphlets about resources. was a small effort and i'm not bragging about it. i bring it up because during my time with that, i'd meet literally recruiting agents. They'd be offering these young kids who were just kicked out of their mom's house a trip to toronto to "work" for them. just street begging where its more profitable. and away from support networks. happens all the time

they've got scouts going all over the country

again, long story short, don't underestimate the capacity of people to abuse good will

[uncut fiber]

tried to see latent space, and it seems i broke my model 🙂 something should be hidden probably. Seems architecture first.
Oh got it, it is latent SPACE, therefore architecture.

[uncut fiber]

cats are my favorite! 🙂

[native zephyr]

nice

[indigo carbon]

@visual glade did UNet2DConditionModel added to import diffusers resolve the BS issue? I just did and it seems to do something

[hardy cipher]

maybe. but they definitely don't look homeless. I guess that's not saying much though. most homeless people really don't look homeless. it's just the shameless and mentally ill ones that we tend to recognize as homeless

[hard fractal]

🅰️ or 🅱️ ?

[hardy cipher]

B

[thorny frost]

B

[crisp owl]

freckle dots to me look too fake in B

[thorny frost]

People have all kind of shit in the skin. A seems too plastic

[hardy cipher]

it's hard to decide with that stuff for me. AI has goofed my perception, lol. real life sometimes doesn't have the same realism as AI

[hard fractal]

Another seed:

[steel mist]

Tbh if I were doing a final, I’d throw them both into photoshop and do a manual blend

[hardy cipher]

B seems like it has details turned up

[thorny frost]

A looks like a Terminator

[steel mist]

A has some “too perfect” elements, B looks overcooked

[hardy cipher]

and I was going to say the same thing as xyrrus

I'd blend them if I was going to be meticulous about it

[steel mist]

One day I will get off my ass and learn how to build a frequency separation node using image magic and python

[thorny frost]

blending in photoshop, applying a mask and editing it

that is the trick

[hardy cipher]

hmm, what would a frequency seperation node entail?

[thorny frost]

you keep the parts you like and blend or discard the rest

[steel mist]

the TLDR is it’s a photo editing technique that tries to separate out fine detail from broad detail. So when retuouching a photo, you tend to do it, then do your editing just on the fine detail to remove things like blemishes

[hardy cipher]

hmm. I wonder if you could do that with existing nodes?

[steel mist]

There’s some convolution and blurring that need to be done

[thorny frost]

editing the mask you mean?

[steel mist]

Sec let me find an article that’s decent so I’m not talking out of my ass

[hardy cipher]

I'll take two images that are close, use the "difference" filter and then use that result to blend back in

[crisp owl]

Is there a setting to tone down B slightly?

[steel mist]

https://www.adobe.com/products/photoshop/frequency-separation.html

[uncut fiber]

isnt it part of krita, wavelet things?

[indigo carbon]

been like this for a while. still happy no errors =]

[thorny frost]

problem with that is there is not a universal solution, each image is different

[steel mist]

We’d basically want to do the opposite though. Retouchers use it to remove detail (blemishes) w/o making an image look edited.

We’d want to add detail by amplifying texture

[indigo carbon]

I'm currently compiling batch 1-16 res 64-2048

[hardy cipher]

I've used the frequency separator in audition

reverse polarity for different pitches

makes things interesting

[steel mist]

Anyways I am not sure how much futzing would need to be done with the node but it seems possible

An in fact the exact sort of thing comfy was built to do

But my coding skillz are, uh, rusty

[hardy cipher]

same. well I've never really programmed in python before recently. but getting acclimated. really just need to figure out how the photoshop thing works exacrtly. then it wouldn't be that hard

[steel mist]

I also have no idea if it would work. Been meaning to overcook an image and then see if I can create a photoshop action

If I can do that, easy to break down and see what’s available w/ existing nodes

But if you can do it in photoshop it’s pretty likely you can just string together a bunch of imagemagick commands

[hardy cipher]

I'm sure they're using some fancy algorithms though

that might be beyond what I could even really understand

[steel mist]

Photoshop has a lot of new hotness that’s proprietary

But all their basic shit is just math

[thorny frost]

Would be nice to have nodes for being able to quick retouch images manually in Comfy

[uncut fiber]

Free Krita got it as well

[cloud dove]

Are there good models for Space stuff?

[uncut fiber]

@cloud dove TDG8UU making nice pictures with copaxTimelessxlSDXL1_colorfulV2
But others can you suggest better models

[cloud dove]

okay thx

[indigo carbon]

it's still going.. did I actually manage to resolve this lol

if this works- I might be the one providing the precompiled modules lol

[half ivy]

guis i suck at sdxl
sad boys club
heart broke island

noooooooot

[crisp owl]

talk to it like it's your friend

whisper sweet nothings

[uncut fiber]

I like this, not finished images, but helping imagination

Latent space is emerging real life world. BREAK
photo-realistic.

[uncut fiber]

would love to have a bit of ice there 🙂

[crisp owl]

just seeing it's a guy surfing/boarding down that haha

[indigo carbon]

CCXL used to be my personal favorite. today I discovered copax_colorful and I like it kinda better

[strong field]

ill take freebies, and wont complain

[hardy cipher]

Very ice

[uncut fiber]

contrast to ideal icy world. Bit of reality

[cloud dove]

how do you upscale ?

[cloud dove]

Damn

[indigo carbon]

it managed to compile something called UNet2DConditionModel.dll which I'm not sure is the right component- but it is compiled for bs1-bs16, 64-2048 res

Idk how to infer it though

hate .json

[hardy cipher]

why?

[terse pewter]

Hi, why cant we enter long prompts in the bot?

Is there any way to enter long prompts?

this is the prompt "concept art for a research laboratory attached within a modern facility, wooden furniture, intricate laboratory equipment, surrounded by a snowy North Pole landscape, glass apparatus in the center table with shimmering green liquid, computer workstations, files on the shelf, 8K concept art, Octane Render and Cinematic Lighting, "

[hardy cipher]

can you not download a user interface?

or use colab?

[steady grove]

ever read the book sphere? one of crichton's best from before he went mental. what if we showed the latent space to that device? oooOoo

movie was okay too i guess

[hardy cipher]

really like these, man. I've been making blueprinty sorts of things latelt

[uncut fiber]

author of that rat?

[steady grove]

jurassic park

[uncut fiber]

o.k. rat out of stainless steel will be somebody else, going to check it

Harry Harrison

[steady grove]

strong name

[stone fossil]

Aye think its kick ass.

[uncut fiber]

the middle one i should keep for some competition 🙂

[hardy cipher]

started off with regular blueprints from google image search. then used those to make weird things with clip vision

[stone fossil]

I made a lora out of it for XL so it helps a lot.

[hardy cipher]

well the blueprint is from clip vision

then that second one is controlnet

I've been thinking about making a lor with them myself

[stone fossil]

It is fun to play with any concept tbh. 🙂

[hardy cipher]

but haven't really done the work to sort through and figure out what I'd like to use

[stone fossil]

SD is one big magic box.

[agile jungle]

am i supposed to generate my image then inpaint it using the refiner model?

[hardy cipher]

that's a new approach but if you can do it I support it

[agile jungle]

what's the standard approach

[steady grove]

i've got a sudden urge to test, "blueprints for a human" but i'm scared to go over that edge

[crisp owl]

https://tenor.com/view/good-morning-do-it-gif-24435329

[stone fossil]

Oh it will do it.

[uncut fiber]

@steady grove like mentioned Schwarzenegger and elephant in one prompt?

@vital ermine Hi!

[hardy cipher]

man, depends on what you use. depends on what you want. you really don't need refiner. you can run it for 20 percent of the steps at the end. no one actually knows how it works. it's a big mystery

[vital ermine]

Hello. 🙂

[steady grove]

but!

[agile jungle]

how do I do that?

run it for 20 percent of the steps

[hardy cipher]

well, you run the base model for a number of steps, then you divide that number by 4 and run the refiner over the image

[stone fossil]

It all depends on what parts u need.

[steady grove]

green print

[agile jungle]

Sorry Im new to this, how exactly do I "run" the refiner? Do I change to inpainting tab and mask the whole picture or something?

[hardy cipher]

peak performance

[uncut fiber]

I really like this style

[hardy cipher]

you mean my wish.com carrottop?

[steady grove]

oh wow til rip roop

[vital ermine]

Yeah, it came out well

[hardy cipher]

I'd suggest looking at some tutorials and things. I'm more than happy to help people. but it'll be complicated for both of us if you don't have some of the base knowledge required to understand what we're talking about. I'm assuming you're using a1111? first tab is text to image, second tab is image on image. not exactly inpainting, although that's an option. you create the image, then send it to image to image/image on image, whatever that's called, and you run the refiner for the final 20 percent of the steps

[hardy cipher]

or if there's a more streamlined approach now, I haven't heard about it

[vital ermine]

hands, the bane of all

[hardy cipher]

whoa, those are nice

[agile jungle]

alright

[hardy cipher]

but yeah, not trying to be dismissive. I was there too, it's a learning curve. but you'll be a lot better off if you read about it a bit

and do some experimenting. if the result isn't what you want that's fine

I think it can be daunting for new people because they see these ridiculous creations made by more experienced users and they compare their results to that

[hardy cipher]

but thing is most of what almost everyone makes is either poo or so so

[hardy cipher]

then a few gems in the mix

personally I like poo gems

I wonder if stable diffusion knows who daddy long neck is

[vital ermine]

well

[hardy cipher]

that's him without the suit on

[vital ermine]

hehehe

[hardy cipher]

wow

a bit harsher than I was expecting

[hardy cipher]

robots living mundane human lives

look at that guy's little sliver teeth

[long vector]

I'm super eager to get Revision working on Colab. I'm following this vid: https://www.youtube.com/watch?v=1SwDCqgXZ-M, but as a nube with ComfyUI, the first thing I'm stuck on: in the vid he describes getting a specific CLIP model and placing it inside the model folder (in a local installation). How does one do the equivalent when running on colab? Alternatively, is there a comfy colab that is ready to go with all the necessary models for Revision?

[hardy cipher]

so revision is just clip vision,, right? or is there more to it that I'm missing?

I can put together a simple workflow with clip vision in a couple minutes. but not sure what else they might be doing in there

if you used a1111 it's a similar model concept. put it in the model folder

[long vector]

thanks very much for your reply: haven't used a1111. Should I place it in the standard colab AI/models path?

[hardy cipher]

you're using colab for it? well it's the same concept. find your comfy folder on your google drive and put the model in there. now if you're not using google drive either. I really don't know.

[hardy cipher]

but yes, standard model folder. I think there's a specific clip vision folder

[long vector]

ok, I see it now on the drive, thanks very much for the help!

[hardy cipher]

https://huggingface.co/comfyanonymous/clip_vision_g/tree/main

put that in there

[hardy cipher]

microwave houses

[wet nacelle]

These look great

[vital ermine]

Yeah, can you imagine living n one?

[hardy cipher]

I'd be alright with it. but I'd always feel like some creeper wsa watching me from below

[vital ermine]

yeah

but fuck having to wash them every single day cause you just know how dirty they would get.

Need a crane for it

[steady grove]

future electrostatic windows that repell materials when charged. and can become opaque. perhaps even stenciled

[ionic gulch]

new version of my workflow online: https://github.com/JPS-GER/JPS-ComfyUI-Workflows
https://user-images.githubusercontent.com/142158778/262469111-88ff5603-5b51-4668-abed-9aefc6b42f08.png

[upbeat summit]

Thanks for sharing it, JPS!

[zinc cargo]

so everyone here still uses refiner?

[long vector]

So I have the model in the specified folder, and the LOAD_CLIP_VERSION node sees it. I've got the graph all connected up as per the vid, but when I try to run, I get this error: Error occurred when executing CLIPVisionLoader:

invalid load key, 'v'.

Does this ring a bell?

According to my-debugging-partner-GPT,

From the provided traceback, it appears that the problem arises when trying to load a PyTorch model checkpoint using torch.load(). The error originates from this line:

pl_sd = torch.load(ckpt, map_location=device, pickle_module=comfy.checkpoint_pickle)

Here, ckpt is the path to the file that's causing the problem. The error suggests that when trying to unpickle the file using pickle_module.load(f, **pickle_load_args), an "invalid load key, 'v'" error occurs.

The immediate file in question is referred to by the ckpt variable in the function load_torch_file inside the file /content/drive/MyDrive/ComfyUI/comfy/utils.py.

[ionic gulch]

my workflow creates four variants with 70-100% base, so if refiner doesn't help you can pick the 100% base version. for many styles refiner helps IMO.

[zinc cargo]

you workflow looks amazing and scray at the same time 🙂

[ionic gulch]

it's easy to use. everything you need on one screen - here is a simple txt2img with little bit of canny for example - only the pose is taken from the voodoo girl, the rest is txt prompt only:
https://i.imgur.com/nUW8R7v.jpg

[zinc cargo]

i enjoyed canny today 🙂

[crisp owl]

I still use the refiner.
I did a test of 100 images with and 100 without and based my own personal feelings on those, comparing each simultaneously. In the end I felt there were better results using it still. Regardless the checkpoint used.

[zinc cargo]

i need to start keeping thing tidy in comfy

[proper depot]

I'm obsessive about keeping my lines clean 😅

[ionic gulch]

i use pipelines on top and bottom to keep it "clean":
https://i.imgur.com/m2xQ8yK.png

[proper depot]

That's hot

[zinc cargo]

dont show us your factorio screenshots!

[ionic gulch]

was better in the past, but i had to add the revision featre, that messed it up a little bit

[zinc cargo]

both 2 and 2.4 does not load for me.
also the png from github

[ionic gulch]

@zinc cargo without any error message?

[zinc cargo]

yes

also not in the terminal

[ionic gulch]

try this one:

[zinc cargo]

works

manger will be able to find all of the missing ones?

[ionic gulch]

did you download the raw file from github? it's a bit tricky do download files from github. you have to right click "raw" and choose "download file form destination" or something like that

[zinc cargo]

oh, i just right click and save

[upbeat summit]

the refiner is a very tricky thing.

• the concept sounds really good and there are some great use cases for it
• it depends on the prompt, because it is heavily style dependent
• it depends how many steps you let it run, because it should be used as a finalizer. so I would never use it for more than 30-40% of the image. For a 40 steps image I let it run for a maximum of 5-7 steps.
• it depends on the sampler you are using
• I have used it a lot and I think it can help in certain situations. I try the refiner in experiments often.
• but almost all images I've posted since the before the launch of SDXL 1.0 don't use it

[ionic gulch]

yes, my own nodes are in manager and the other ones are from very popular node packs.

@zinc cargo someone had problems with one node from impact pack yesterday. please try it if it works for you

[zinc cargo]

maybe it's not my day today

::\

[ionic gulch]

did "find missing nodes" in manager work in the past for you?

[zinc cargo]

yes

[zinc cargo]

yes

i'll do an update just in case

usually solve stuff

[ionic gulch]

you can find my own nodes if you search for "jps" in manager

[zinc cargo]

if i would ever made something for comfy, it would be to restart from withtin the gui

[hardy cipher]

how did it go when you submitted them?

I mean, obviously they're there now, lol

[ionic gulch]

the rest are common packs like

• ComfyUI Impact Pack
• (WIP) ComfyUI's ControlNet preprocessors auxiliary models
• WAS Node Suite
• Comfyroll
• Derfuu

[ionic gulch]

i made a fork of the manager, changed the file with the sources and made a pull request.

[hardy cipher]

ahh. I was reading about it. seems simple enough. just hadn't used github before now so new concepts to me

well I'd downloaded things from it, but never really interacted

[ionic gulch]

was my first time too. not too hard. took me 10-20 minutes of google/youtube to learn how to fork and make a pull request.

[hardy cipher]

lol, DO NOT ask an llm how to do any of these things

fyi

not sure if they're getting "dumber" or if I'm just used to them now

[ionic gulch]

i would explain it to you, but as i'm a n00b github user myself, i guess a youtube video will be better than what i remember from the process :)

[hardy cipher]

ahh, nah, I think I have it. it's not that complicated. I just get anxious about the idea of missing things

[visual glade]

https://huggingface.co/TencentARC/T2I-Adapter/tree/main/models_XL
These are now supported by ComfyUI

[hardy cipher]

how do they differ from the ones that stability released?

[zinc cargo]

which one is CR? the comfyroll one?

[hardy cipher]

never quite figured out the differencece with the t2i stuff

[hardy cipher]

I'm pretty sure

did you try the missing nodes button?

[ionic gulch]

https://i.imgur.com/h4JJp8w.png

[upbeat summit]

https://github.com/RockOfFire/ComfyUI_Comfyroll_CustomNodes

[zinc cargo]

yes, it doesnt work for me.
the comfy roll is installed, but doesnt load

[hardy cipher]

lol, how is the code indented incorrectly?

[upbeat summit]

this doesn't look right

I think there's some other conflict

[hardy cipher]

that would make more sense

[upbeat summit]

I just had a custom_node that kept UltimateUpscaler from loading

so conflicts do exist

[hardy cipher]

I had an error last night that caused me to end up deleting a few big node packs

turned out it was something entirely unrelated that wasn't showing up in any error

well, it was related. wrong word I guess

just hiding

[zinc cargo]

how did you find out which one was it?
both for @upbeat summit and @hardy cipher

[hardy cipher]

pooled output was my error

[ionic gulch]

here is a list of nodes i'm using without conflict at the moment:
https://i.imgur.com/SrD7Ww4.png
https://i.imgur.com/QUy12VY.png

[hardy cipher]

well, need to see more of the error tbh

I just thought of what new nodes I was using. not new node packs, but specifics I hadn't used much before

[upbeat summit]

so first I check what's up in the console. but for the UltimateUpscaler thing the node that was responsible didn't even throw an error. so I just removed the node packages one by one, till I found the one causing the problem - or just start from scratch and copy each one back one by one

[hardy cipher]

becaause I'd get the errors, stop getting the error. then it'd reaappear when doing something seemingly simple

last line got cut off in that screenshot, but that was pretty much it

pretty obvious it has something to do with the sampler, or something that had connected to the sampler, so just kind of had to do deduction work

[zinc cargo]

no stability nodes?

[hardy cipher]

simplest solution normally is to update everything

[zinc cargo]

is there like a git . i can run recursvly ?

[ionic gulch]

no, i don't have them installed on my system.

[hardy cipher]

for what?

are you using a virtual environment? if you want to be a mad lad you could run this in your comfy folder in powershell

Get-ChildItem -Recurse -Filter 'requirements.txt' | ForEach-Object { pip install -r $_.FullName }

well I only did that because something got goofed and a bunch of python packages didn't get installed

[zinc cargo]

even matching your nodes, i get the same thing

[hardy cipher]

and everything is updated?

[zinc cargo]

yes

[hardy cipher]

have you tried just restarting and refreshing the browser?

[upbeat summit]

it really does help 😄

[zinc cargo]

the error is at the server level

[upbeat summit]

what server?

[zinc cargo]

i know, i'm that guy who everyone calls to help about their pc as well 🙂

[ionic gulch]

try to remove my pack (JPS) and check if the others work or show the same error

[upbeat summit]

hah same

[zinc cargo]

the comfy server loads from terminal...

[upbeat summit]

so move all nodes out of the custom_nodes folder

and move them in one by one

find the culprit

[ionic gulch]

@zinc cargo just did an update and get the error now too - so my guess is that the comfyroll-pack is broken at the moment.

[upbeat summit]

that is a possibility too

[zinc cargo]

it works as a standalone

[upbeat summit]

it may conflict with a node we all use

[hardy cipher]

well what node is it?

[upbeat summit]

WAS, efficiency etc

[zinc cargo]

no wait!

File "C:\Users\User\SD\ComfyUI\custom_nodes\ComfyUI_Comfyroll_CustomNodes_init_.py", line 1, in <module>
from .Comfyroll_Nodes import *
File "C:\Users\User\SD\ComfyUI\custom_nodes\ComfyUI_Comfyroll_CustomNodes\Comfyroll_Nodes.py", line 1572
class Comfyroll_LoadAnimationFrames:
IndentationError: expected an indented block after 'else' statement on line 1566

Cannot import C:\Users\User\SD\ComfyUI\custom_nodes\ComfyUI_Comfyroll_CustomNodes module for custom nodes: expected an indented block after 'else' statement on line 1566 (Comfyroll_Nodes.py, line 1572)

[upbeat summit]

yeah

[hardy cipher]

yes, indentation error

[zinc cargo]

still that ident thing... no nodes other folders

[upbeat summit]

so comfyroll broken

[zinc cargo]

should i just ident it?

[hardy cipher]

well indent that mofo!

[zinc cargo]

i mean, it's just pressing tab, right?

[hardy cipher]

I mean, just undo it if it doesn't work out

[crisp owl]

could be as simple as a tab instead of a space

[ionic gulch]

https://i.imgur.com/yF6XVAn.png

[zinc cargo]

https://tenor.com/view/yay-crazy-face-im-going-in-uptown-girls-gif-13802985

[crisp owl]

I made that mistake last night editing something, tabbed, broke it, then realized it expected 4 spaces instead of a tab. easy correction

[hardy cipher]

dude, what kind of psychopath came up with the space/tab thing?

[upbeat summit]

this python thing is crazy

[crisp owl]

in notepad++ there's a make all tabs to spaces button, so that comes in handy

[quartz sequoia]

okay i cant i love the generations im getting from comfy.. but i want Auto gui and features... with comfy workflow backend.. and somehow magically they work cohesively.. comfy generations are better also.. but i just cant node system workflow is a bit messy sometimes

[hardy cipher]

I like that I can also highlight a bunch of stuff in notepat++ and tab it all at once

[quartz sequoia]

stableswarm works until it doesnt with some workflows because idk

[zinc cargo]

here is the line with the eles

[hardy cipher]

actually didn't know about this though

[quartz sequoia]

and comfybox..has potential but the dev left it

[hardy cipher]

what are you talking about?

[crisp owl]

under "edit"

[hardy cipher]

generations aren't different depending on user interface?

[eternal fog]

If that's the code that's erroring it's probably doing it because there's no actionable code in the Else statement. With it commented out, there's no reason for that else to be there.

[quartz sequoia]

comfy generations are different than autos

[zinc cargo]

so i can just commnet it as well?

[quartz sequoia]

also speed is better than auto

[hardy cipher]

so much misinformation

people mistake their opinions for facts

[quartz sequoia]

its not a opinion..

[zinc cargo]

Maki, cant you see i'm in crisis?

😛

[hardy cipher]

alright, explain yourself, maki?

give the details

[upbeat summit]

python needs indentation to work, which is crazy, but this is how it works

[quartz sequoia]

WE are all in crisis mate

[quartz sequoia]

also whats ur issue tho

[upbeat summit]

ComfyRoll will probably be updated in T-X min 😄

[eternal fog]

Yeah comment out the Else, or just delete it

[quartz sequoia]

comfy with the same settings generates a different image on auto

[upbeat summit]

ComfyRoll updated 2 minutes ago @zinc cargo

https://github.com/RockOfFire/ComfyUI_Comfyroll_CustomNodes/commit/438d9ae866d525ab5133ecac19b016be5a634f7d

[hardy cipher]

if it were me I'd just delete the comfyroll install and git clone it again

[hardy cipher]

yeah, because you're probably using a different scheduler or something

[quartz sequoia]

maybe yes and maybe no

[eternal fog]

It's not just that Auto does some different stuff with the images that it doesn't present to the user. There is no way to get identical images between Auto and Comfy

[hardy cipher]

maybe you're assuming things when you're not sure about them. that's all I'm saying

[zinc cargo]

https://tenor.com/view/high-fiving-high-five-gimme-some-skin-greeting-congratulations-gif-14414580

i'm such a haxxor

[quartz sequoia]

im assuming my assumitions are correct

which im pretty sure they are

[hardy cipher]

okay, I've seen people say so many things like this, and then they're not accounting for something simple

[quartz sequoia]

did you try to chatgpt it?

[hardy cipher]

he fixed it

[eternal fog]

Comfy has said something about it before. I don't remember the details, but there's something that Auto's code is doing differently, which means it's basically impossible to get the same images, even if you use the exact same settings between the two.

[quartz sequoia]

not a fan of accounting ngl..

[hardy cipher]

alright, you've deflected a few times meaning you probably don't have the details of what you're doing

which is fine

[quartz sequoia]

i remember it was that comfy generates noise using the cpu somefing like that

[hardy cipher]

I just wish people wouldn't spread misinformation

[quartz sequoia]

while auto idk does something else

[ionic gulch]

this is the fix:
https://i.imgur.com/f9aIuZg.png

[eternal fog]

The fix would be deleting those lines

[ionic gulch]

add # in front of else

[zinc cargo]

thanks i did exactly that!
so proud of myself

[crisp owl]

it's been patched in the github

removed those lines completely

[zinc cargo]

and now, to generate some stuff

[eternal fog]

There's no point in commenting out lines that don't do anything, just delete them

He's clearly put that in for debugging and forgotten to remove them

[hardy cipher]

it's already hard enough for people to understand this stuff. don't need to have other people going around planting little seeds of half truth

[upbeat summit]

yes it is 😄 https://github.com/RockOfFire/ComfyUI_Comfyroll_CustomNodes/commit/438d9ae866d525ab5133ecac19b016be5a634f7d

[zinc cargo]

but before that...

[crisp owl]

heh sdxl makes fun creatures

[hardy cipher]

they look kind

[glad grove]

mike wazowski

[ionic gulch]

doesn't really mater for a quick local fix while waiting for an update. but it seems the update is already available, so updating through manager is the real fix :)

[crisp owl]

I did some as well the other day

[eternal fog]

Yeah it just causes problems, like the one you were trying to fix.

I know I've done it before

[hardy cipher]

I like how in each one he looks like he's really holding in a poo

[glad grove]

holdin diarrhea lora

[ionic gulch]

@zinc cargo so my workflow now works for you?

[zinc cargo]

i just hit ctrl enter for the first time

[upbeat summit]

https://tenor.com/view/keyboard-rage-keyboard-a1esports-shred-keyboard-break-keyboard-gif-22428093

[long vector]

That did the trick, thanks! Now that I have it working, it is interesting to compare to image prompting with MJ and also Kandinsky. Of the three methods, Kandinsky comes the closest to matching the features of both input images, and not adding what seems like its own spin. For example, when I used two b&w images with Revision, I received color images as output, something that never happens with Kandinsky. I'm certainly eager to experiment around some more, though.

[crisp owl]

COuld be as simple as the weights for them

[zinc cargo]

so trye bcz i have 0 idea what goes on in this glorious beast of yours

[ionic gulch]

txt2img + canny with 70 steps takes 42 seconds on my system for the 4 versions:
https://i.imgur.com/qbShbRC.jpg

[indigo carbon]

working on batch size for AIT, base works- now compiling refiner

[crisp owl]

And then sm75 support!

[indigo carbon]

don't have SM75 to compile on. FizzleDorf might though

they said they will add easy compiling scripts to the node's files so everyone can get the highest boost possible

[crisp owl]

What's the compile process? Is that the only thing holding back currently sm75 support?

[ionic gulch]

canny with low value is nice to get a girly dog:
https://i.imgur.com/fJoEsXs.jpg

[indigo carbon]

yes. it's the process that AIT does when it makes a specialized module.

theoretically, it can all be precompiled.. but that's a job for someone like Comfy that has access to all that

[crisp owl]

Ah ok, so nothing i can really do on my end

[indigo carbon]

you can, compile yourself using AIT

that will become easier eventually

[ionic gulch]

or scifi dog:
https://i.imgur.com/zZDmrWT.jpg

[crisp owl]

Yeah Unless there was something that specified what needed to be done, that seems out of my current realm

[ionic gulch]

or house wife style:
https://i.imgur.com/KTqoORK.jpg

or spider-alien-dog:
https://i.imgur.com/Ourqdty.jpg

[zinc cargo]

sorry, i still dont really get the workflow 🙂

@ionic gulch best so far

[ionic gulch]

already looks good. if you have questions just ask. basically you have to switch between txt promt or img promt and one of the four generation modes

[zinc cargo]

@JPS i tried with canny, to make the ones above...
so i chose 3, made but i'm not sure if which one should i put the input of the canny to. img2img or revision?
does revision keeps working for canny?

[ionic gulch]

the revision positive img replaces the text prompts completly if you switch revmode to IMG promt, if you use TXT Promt the image is ignored and the text prompt is used.
for canny, depth, img2img you use the img2img image and it shows you a preview of the canny, depth mask on the bottom

so the left image is only for the promt mode and the right image for the generation mode

should make the revision node yellow i guess, so it's more clear

[zinc cargo]

what about canny settings?

also sorry about these 🙂

[ionic gulch]

canny settings are in the generation mode node (blue):
https://i.imgur.com/ZXfuc4r.png
strengh + 4 ctrl options. i mostly use strength and ctrl_stop

i've changed the colors and titles of the nodes to make it more clear:
https://i.imgur.com/D559oPm.png

[zinc cargo]

is this img2img?

[wet nacelle]

naw

[wet nacelle]

This is all from these images. @zinc cargo

[high skiff]

@visual glade I think I might have found a code issue in comfy UI, but I am looking into it. Running some huge tests for my 2.0 high res fix. I am starting at step 296/300, going to step 300

That should be only 4 steps, yet its queuing 40 of them, and its also slowing down considerably over normal

or sorry, 294/300

so 40 instead of 6

[indigo carbon]

DONE

[high skiff]

ok, its an error with setting the end step above 300

I have it set to 1000 to stop at, which causes it to do 40 steps instead of 6. So weird

[indigo carbon]

refiner is compiled for BS1-4

[indigo carbon]

res 64-2048, this module should be universal for any gen params as long as you're on windows with SM80

[zinc cargo]

and text prompt ?

[ionic gulch]

i think there is a rounding problem with ksampler and high step values. i've reported it here:
#✨｜sdxl message

[wet nacelle]

yeah.

blue and pink candy neko boy wearing maid outfit at a bar, by hiro mashima, akira toriyama, Masashi Kishimoto

[high skiff]

interesting, thanks!

also, little tease for the public

I redid some settings on my v2 high res fix, and it now pulls out even MORE texture and detail over my previous 2.0 candidate

Left is base SDXL, middle is my old 1.0 high res fix, and right is my new 2.0 candidate

[ionic gulch]

here is another example:
#✨｜sdxl message

[glad grove]

this one is best

[wet nacelle]

gold

[long vector]

Yes, I did try playing with the weights, still getting color, and also imagery that isn't in either of the two sources. The strange thing is that both are b&w images. I haven't taken the time to try to understand the flow of the graph yet, but it looks sort of what I'd imagine you'd get if you used Clip Interrogator with both images and then used those descriptions as text prompts, whereas with Kandinsky it is very close to the style of the source images.

[crisp owl]

Ah interesting.
Yeah I've only done a handful of images through the new cvision method, currently trying to resolve another issue, if possible.

[zinc cargo]

@ionic gulch thanks!

[high skiff]

Another comparison

base SDXL, old high res fix, 2.0 candidate for my 1.1 workflow

[wet nacelle]

@hardy cipher

[long vector]

I'm curious how the technique for image prompting in Kandinsky differs from this sdxl workflow.

[crisp owl]

Beats me. I've used A1111 and now Comfy

[autumn forum]

This looks like my Cinemax model. Is that the model you’ve been using for you pics? Love your generations btw.

[steady grove]

kadinsky is a whole entirely different model architecture from stable diffusion

like dall-e

[wet nacelle]

yeah it's just the one I've always been using for hyper realism. I can get better texture detail if I weren't using these images or this workflow, but without these images I wouldn't get this style either.

[autumn forum]

Ooo nice. The model does need some help, hopefully I can get it lookin better soon

[cedar tusk]

Does anyone making ComfyUI nodes have an idea of how to get a custom node to show an image preview like the LoadImage default node does? I fear that the feature is hard-coded based on node name...

[west breach]

you have to do something like this

[cedar tusk]

hmm

[autumn forum]

That’s a cool idea like preview preprocessors or preview image upscale basically preview anything. Rn it seems to be just samplers

[cedar tusk]

I know that works for output nodes like image saving

Does it also work for input nodes*?

[high skiff]

Dialing in my new highres fix even further

Left is base SDXL, middle is my old high res fix, right is my new textured high res fix

[high skiff]

(Please excuse the weird right band on my comparison images there, its a funky seed issue)

[west breach]

not sure sorry, haven't made a node like that

[crisp owl]

was curious what that was lol

[high skiff]

Yeah, it seems to be an issue on some seeds using my LoRA

[west breach]

Think I need another coffee

[high skiff]

evidently, one of the images in my dataset seems to be introducing it

[crisp owl]

dang

[high skiff]

which is weird, cause they are all super clean images, and I only seem to get it with this prompt

I have seen it a couple times

[indigo carbon]

got even faster.. batch size>1 support soon =]

[cedar tusk]

If I just outright copy the LoadImage node from the base code but rename it to something else, no image preview and no upload button 😦

I think it is hard coded

[west breach]

this didn't work?

[cedar tusk]

Because it's an input node, it's returning other things (image, mask, and a bunch of metadata)

[visual glade]

give me a workflow if you think you found a bug

[cedar tusk]

Oh, hey, there's someone who probably knows!

[high skiff]

will do, just need a bit

[cedar tusk]

@visual glade Is the ability of a node to display an image preview hard coded to the LoadImage and LoadImageMask nodes? I have a custom node for loading images but it doesn't show a preview

[half ivy]

@glad grove

https://youtu.be/3v4-zYghodc

[glad grove]

sensei 🙇‍♂️

[indigo carbon]

Also, @visual glade , the fix was adding import UNet2DConditionModel to the beginning of the code. Just compiled for XL and XLR on SM80 with bs1-4 res64-2048.

[hardy cipher]

@cedar tusk Hey, had a question about one of your nodes. Specifically the positive-negative with text node

[cedar tusk]

sure

[hardy cipher]

For some reason, I got this error when using it last night. Have you seen it before? Last line was cut off, but didn't have anything specific in it

[cedar tusk]

let me check it out, I haven't used that node in a while (I made it before I was aware of primatives and being able to change toggle inputs and widgets, haha)

[hardy cipher]

The node wasn't listed in the error, but figured out it was at least facilitating it

[visual glade]

Here: https://github.com/comfyanonymous/ComfyUI/commit/d7b3b0f8c11c6261d0d8b859ea98f2d818b7e67d

[cedar tusk]

Great, thanks!

[vital ermine]

A little more ominous

[cedar tusk]

I just tested a basic workflow using that node and it seems to work. I'm seeing no error pop up BUT I did just find a seemingly different bug with the filename using timestamps...

But anyhow, if you remove that node from your workflow does it still error out?

Also, are you using SDXL with it? I think that SDXL requires a different CLIP text encode node due to internal differences, right?

[hardy cipher]

Only got that error with that node. Never saw it otherwise. And yeah, used xl with it. Not a big deal, just thought I'd bring it up with you

[cedar tusk]

Ah, I'll see to spooling up an XL version then

[vital ermine]

Dude been on many a trip with no spaceship needed.

[steady grove]

these are the voyages of the starship d o double g

[cedar tusk]

@hardy cipher So what do you think about the positive_text returning combined pos_g and pos_l?

[vagrant helm]

https://minimaxir.com/2023/08/stable-diffusion-xl-wrong/

[cedar tusk]

neat

[vagrant helm]

Sorry if this was already posted. It's pretty interesting. They intentionally generated a bunch of crappy SDXL examples, then trained a LORA with those to use as a negative prompt and it works better than you'd expect

I thought the base model was already RLHF'd but maybe it didn't get enough negative reinforcement

[cedar tusk]

@hardy cipher Okay, SDXL version of the CLIP Positive-Negative nodes are now up

[native knot]

Get that chedda

[hoary saddle]

anyone know what "ImageScaleToTotalPixels" node comes from?

lol

[steel mist]

hrm no safetensors version of it though 😭

[hardy cipher]

Sounds decent to me. I've been exploring different options in that regard

[vital ermine]

LOL

I think this locon was a success.

although lycoris locon has a bug now so I had to use kohya

[hardy cipher]

Damn, those are legit. What did you use for training?

[vital ermine]

200 images from a movie

then V2 was 150 and finally V3 120

[hoary saddle]

anyone know where this comes from?

[west breach]

comfy_extras

[vital ermine]

my problem is I can't think of any prompts to try it on

the lexica prompts are meh

[ivory blaze]

@vital ermine I just start writing something from a small base idea and add random stuff to it as I move through and just switch subject matter here's a prompt for you: a 1970s album cover in the style of grand funk railroad featuring a photorealistic tyrannosaurus rex playing golf with John Lennon in the landing bay of a large space station orbiting a gas giant

or you could try to get it to create a photorealistic image of a scuba diver wearing a scuba tank... which 1.5 nor any refined 1.5 model I have tried can, I know cannot do at all for some reason and others tried sdxl earlier I think, and said similar problem lol

[elfin flare]

Beautiful.

That is an IMG2IMG of my cat.

[ivory blaze]

here's an img2img upscale of whatever prompt I use, I forget

actually that may be a 1.5 i need to organize this stuff

here's a few I made earlier with sdxl

[ivory blaze]

that one is 8k though :\

[vast galleon]

I've successfully integrated 3 new color transfer methods into a custom node. Now to incorporate the old method on the bottom into it.

The big picture is to get GroundedSAM in as well to segment out masks and edit masked regions with color transfers from different images

[tepid surge]

what a long road

[vast galleon]

Once GroundedSAM is in, this could possibly be added for automatic openpose MoCap

https://arxiv.org/pdf/2303.16160.pdf

https://github.com/IDEA-Research/OSX

[tepid surge]

I just want what I know is mine