✨｜sdxl

[soft zealot]

you can use either , its just you may need a different workflow for Dreamshapper 8 as its a Stable Diffusion 1.5 based model not an SDXL based model.

there is a link on their Civitai page to the SDXL version of Dreamshaper, would you like me to get it for you?

[next night]

I'd like to but dream shaper 8 is bad, lemme show you

[soft zealot]

jeeez lousie , if you are using a SDXL workflow and Dreamshaper 8 I am not surprised its "bad".

As syated Dreamshaper 8 is *NIOT an SDXL based model

[native knot]

Dreamshaper 8 is fine if you adjust settings. It's a very popular checkpoint for 1.5.

[uncut fiber]

still asking for how to download extensions in Comfui?

dreamshaper and deliberate are great

[soft zealot]

depends meither follow the instrucstion on the maintainers guthub|civitais|hugging face page or use ComfyUI Manager

[native knot]

ComfyUI Manager is great.

[cedar lark]

been getting error using my controlnet workflow since the update, normal gen works just fine, the error attached

[native knot]

Makes grabbing missing extensions easy.

[uncut fiber]

i know, i thought if jinglebels still asking. @soft zealot thank you very much. I have Manager 🙂
Sorry for confusing you

[soft zealot]

@native knot hows this , small tweak to the prompt plus its now Dynavision for Primary and talmendox as secondary

yes I know , brown hair shaped into ears but it actually makes sense lol

[native knot]

That's awesome. Though it still has the ear issue. 😄 😄

That look is great, though.

Same seed as before?

[soft zealot]

no

[native knot]

Ok, was going to say...

[soft zealot]

but I can go back

[native knot]

Did this last night...Tom Hiddleston as Joker:

[cedar lark]

just FYI, the workflow I am using is attached here, I have not modified or changed it since last night, and it worked fine yesterday

[soft zealot]

looks at image name 219042278286667 is the one I want

[soft zealot]

Ok same seed as the original 2

Just goes to show, so many variables and that at the end of the day a lot of it is just luck. We are putting random words in a pot, sprinkling in some randow numbers, giving it a mix and oping something tasty pops out

[native knot]

Interesting take with the shorts. 🤣 Definitley better with the other seed. Still really a great look overall, though.

I'm working on a Catwoman, currently, but I have a meeting starting in a minute. 😦

[strange mist]

show up late with a good catwoman image. I'm sure everyone will understand 🐱

[soft zealot]

@native knot , here we go have some black ears 🙂

[native knot]

Getting closer for sure.

[strange mist]

only missing the cape now 😮

[soft zealot]

one moment caller

hmm curious blend between boots & legs

and the face has drifted away a little by adding in the cape

[cedar lark]

just to update on this: apparently using the "simple" controlnet module works, but using advanced controlnet with start and end percentages fails, so it's to do with that

[next night]

Look at the difference between mine and the owner's, i used same prompts and generation data and same model@soft zealot @native knot

[potent flame]

i mean it is pretty ass for me to use my 2GB VRAM for SDXL generation

smh A1111 run faster than ComfyUI

[strange mist]

other than the legs, the image is pretty good 🙂

[potent flame]

in most occassion

[next night]

Bruh im using 16vram and its bad i think im doing something wrong

[potent flame]

it is more like bicycle-ride-on-the-highway situation when you can have OOM almost every single second

[soft zealot]

sticking with the same seed for now, tweak one thing and something else drifts lol

[native knot]

Yeah...I'm always battling that.

[potent flame]

xformers dont run for me that well that most people will expected to in their 3060

[soft zealot]

no offenc but 2Gb I wouldnt even waste my time.

[potent flame]

it is truly 4-6 min per pic in 1.5

you know how much does it take to generate 1024x in SDXL?

especially in ComfyUI?

[strange mist]

love the thigh high boots 😄

[next night]

I have rtx 3060 16GB vram and 5-7 seconds for an image

[potent flame]

basically it will just take 40 min per pic in SDXL

[soft zealot]

I can only speak for myself. I run a 1080Ri 11Gb and a basic 1024x1024 SDXL based image takes approx 50-65 seconds to generate in ComfyUI depending on sampler & actual model etc

[potent flame]

yeah most people dont have much patience.

[next night]

Is that alot or normal

[potent flame]

but it been quite a ride with my bicycle

[wet nacelle]

https://tenor.com/view/jerma-jerma985-a-mimir-a-mimir-meme-mimir-gif-24965637

[potent flame]

you know like

when it show up Out of Memory needed 0.5 GB VRAM of space every single time it done generation

[soft zealot]

its about whats expected for the card I have

[potent flame]

after attemping to VAE decoding

yeah I think for a 1080Ti, that is pretty normal

[next night]

What about dream shaper8? How many s it takes

[soft zealot]

6gb Vram is probably the minimum amount of VRAM to use for SDXL based models although I would say 8Gb upwards better

[potent flame]

yep

[wet nacelle]

frens

[potent flame]

technically you can run it on 4GB VRAM thought.

[soft zealot]

no idea, I havent got it and I havent rin any other 1.5 models recently in COmfy but is more like 15-2o seconds IIRC for a 512x512

[potent flame]

but that is bare minimum

[soft zealot]

doesnt mean it will be any good.

Technically you can run it on CPU only and not even have any VRAM, doesnt mean you should!!!

[native knot]

Batman be trollin'.

[hoary saddle]

Nobody else is getting system freezes after a bit of rendering in Comfy?

[strange mist]

super frens

[soft zealot]

@strange mist @native knot think thats got it 🙂

[native knot]

Nice!

[potent flame]

what is the Prompt G and Prompt L?

[native knot]

I may have to toy with that one later a bit.

[potent flame]

like I know this is the new kind of CLIP

[wet nacelle]

https://tenor.com/view/cut-meme-funny-i-like-ya-cut-g-gif-19499891

[strange mist]

nailed it, chief 🙇‍♂️

[soft zealot]

how much system RAM ro you have?
<=16Gb not surprised
<= 32Gb , possibly
I run 64 Gb 🙂

[hoary saddle]

32 with large swap/page

[next night]

Soul why do you sound like a dad

[next night]

By rendering means generating?

[wet nacelle]

@soft zealotHow can you bare using XL the way that you explain? 50 - 60 seconds per gen?

[native knot]

64GB RAM users unite!

[potent flame]

SDXL is RAM-monster basically

like Chrome

[wet nacelle]

bair bare bear

[soft zealot]

do you mean @strange mist rather then @native knot

And whats wrong with Chief?

Come back Guy Martin!!

[soft zealot]

because Im a nearly 60 year old who has all the time inthe world and unlike the pimply faced youth of tday I dont expect instant gratification serverd to me on a silver spoon

[next night]

Bruh what the fart is coming out of u keyboard, stfu 😭😭

Ididnt understand shi

[wet nacelle]

https://tenor.com/view/crying-man-sad-and-depress-gif-13089667

[hoary saddle]

Was kinda hoping the 4090 24gb upgrade wouldn't really need the system ram as much

[wet nacelle]

I see what you mean

[strange mist]

no reason to be disrespectful 😒

[cedar lark]

so just to update more: apparently it has to do with negative prompt conditioning through controlnet module, detaching the negative conditioning into sampler worked, even with the start/end percent "advanced" module

[soft zealot]

May I presume sir/madam (delete as applicable) that you have no conception of who Guy Martin is?

[next night]

Apologize

[next night]

Nah idk who chief is

[wet nacelle]

?

singular?

[soft zealot]

eh? Chied as in "Mate", as in "bruh" , as in "bud" etc tec

Its's a colloquialism

[potent flame]

oh wait

memory management?

for RAM?

[next night]

I thought like the chief who cooks

[soft zealot]

thats a Chef, think Salty Balls

[next night]

Batman is from kryptonite planet

[next night]

Lmao how do you know that balls are salty

[soft zealot]

I've watch south Park. Chef sings about his salty balls

[next night]

Dam

Winston

What does lora means

[soft zealot]

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

[wet nacelle]

Same Mother

WOZA

[next night]

Fix it using inpaint

[strange mist]

marthaaaaaaaaaaaaaaa 😭

[soft zealot]

depends on rontext. It could nmean LOng RANge , TBH I dont know what it actually stands for in cntext of SD

[next night]

No imeant for stable diffusion stuff

[wet nacelle]

Let me spark some conversation here.

This is how I'm getting my styles so quickly.

THis UI is new and made by lllyasviel.

https://github.com/lllyasviel/Fooocus

[soft zealot]

thats what I just daid I didnt know, read the full sentence a

[next night]

Oh

[strange mist]

how do you find it? I heard it's a bit limiting for advanced users? 🙂

[wet nacelle]

this can aso run on 4GB of Vram

[next night]

Mack my auto1111 is stuck on installing gfpgan

[soft zealot]

oh all the Srability AI styles were posted in here a while back in CSV format and there are a number of Nodes that exist to load them

[next night]

Im using comfyui now and its hard

[wet nacelle]

It is. I like it for easy no think shit.

[wet nacelle]

I don't use that anymore. I had no issues with my downloads either. I'm sorry.

[next night]

I wish someone tells me how to fix it

[soft zealot]

Its ideakl if yuore a MidJourney User who wants to stop paying and run locally otherwise its a waste of space IMHO

[wet nacelle]

Oh rad.

[strange mist]

I'm the dude who mostly steals other ppl's comfy workflows 😄
might be perfect for me.

[soft zealot]

I personally use
https://github.com/twri/sdxl_prompt_styler

I have 3 of them that I use in various combinations

[fair stag]

Fyi, something similar to that is built into the dev branch of auto11111

[strange mist]

I should take it out for a spin 🤔

[soft zealot]

feel free to steal mine but make sure ytyou read the Credits & Notes section fully before asking any quaestions ;o)

[wet nacelle]

This UI also has a folder for loras too. This UI is built entirely for XL 1.0 and 0.9 (they are merged into one base and refiner)

It is also fact that you can not replicate these results in any other UI due to how lllyasviel
coded the internals@strange mist.

[strange mist]

I've stolen several versions of your workflows already, but every time I open them and try to understand, I get panic attacks 😁

[soft zealot]

PMSL

[potent flame]

oh wait it is your work flow?

[native knot]

I've used this as a helper for the pixel art quite a bit, actually. Does wonders at times.

[wet nacelle]

This is funny.

[strange mist]

oh, interesting. can it run custom sdxl models?

[wet nacelle]

Yes. I think. I'd look through the Git to see if it's mentioned anywhere.

[strange mist]

thanks. I'd def give it a shot.

[wet nacelle]

I mean this is here but I don't know how much he intendeds his code to mix well with custom models.

legs

[strange mist]

possibly will add support in future versions.

[wet nacelle]

Maybe

[potent flame]

how do you guy use LORA in ComfyUI?

nvm

realizing it will need extra node for that

i guess I could master Blender node after using ComfyUI

[soft zealot]

@strange mist @native knot same Prompt|seed erc but using NightVision+ talmendoxl rather than DynaVision+ talmendoxl

[native knot]

Oh, my. That's nice. I wonder if you'd be able to get the belt and/or bracers back on there without too much issue. It would really pop if that happened.

[soft zealot]

the mask has gone a bit onky as well

[native knot]

It's not terrible, though.

[soft zealot]

This is 100% NightVision

[native knot]

The pose is what I would've expected from that...but the other pose is better and more suited to Batwoman/Batgirl.

[pure hazel]

Honestly, hat's off to the team for SDXL. It's SOOOOO flexible as a base model. It REALLY and I mean REALLY responds to prompt tokens well. I'm blown away by the creativity! I can ask it for almost any ethnicity and it does such a fantastic job. Bravo

[clever verge]

Don't think SDXL has a complete understanding of cat-woman. 😄

[polar jacinth]

Can someone make a person in the world of harry potter realistic, mine only comes out semi-realistic

[soft zealot]

@strange mist

[clever verge]

He's a cool best men(?)

[unreal plover]

Is here a confyUI node that auto select a lora from the prompt ? Like if we set lora:name:1 in the prompt, it will auto load the lora ?

[hoary saddle]

guess i'm gonna wipe clean on one of the systems, dunno what else to try, Ubuntu 22.04 still best option or do we know what linux version all this prefers?

[upbeat summit]

@indigo carbon do you know of any problems when doing portrait images using AIT? Like 840x1176. to me it looks like on the sampler the image is getting diffused in a quad grid, but only 1 part of the grid is getting diffused correctly.

have you seen that before?

[crisp owl]

legion posting nice images and poofing them, teasing us

[upbeat summit]

yeah it's beautiful 😄

[indigo carbon]

nope, however- you might be using an old version of the AIT workflow, which wasn't compatible with any AR. if you are using the latest workflow, idk what's causing that.

[upbeat summit]

only happens in these aspect ratios to me - landscape is fine

hmm interesting

[indigo carbon]

hold up, let me try

[upbeat summit]

thank you

[indigo carbon]

@upbeat summit seed+prompt?

[uncut fiber]

i got such bad image trying to render fullhd.

[upbeat summit]

you can use any prompt, any seed. just try 840x1176

[upbeat summit]

if you work in these out-of-spec resolutions you will have more bad ones than good images. if it works they have amazing detail, but it's a big seed lottery game

[peak dove]

On MJ, if you simply use the prompt 'beauty' it'll always be a woman and always a white woman at that!

[uncut fiber]

o.k. thank you for info, i thought it is imposible. @upbeat summit
try 848x1184 curious if it will change. Because both will be divided by 16.

[upbeat summit]

these are all native 1920x1024 - no upscale

[indigo carbon]

this is very odd, I'm gonna report this

[upbeat summit]

congrats

we found a bug 😄

[indigo carbon]

it's just on that specific res though

[upbeat summit]

must have something to do with the modules I suppose not correctly diffusing that aspect ratio

[clever verge]

The artsy bug

[upbeat summit]

thanks for testing it!

and confirming I'm not crazy 😉

[indigo carbon]

current fix is using a slightly different res that triggers different modules, it appears that whatever modules it's using for that specific resolution are broken

[upbeat summit]

yeah

[vale eagle]

It is actually very cool with that effect

[timid garnet]

.
https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/b17834e9-4894-4caf-ba73-72d26b212e22/width=768,optimized=true/ZavyChromaXL2.jpeg

[uncut fiber]

as i thought 1064x1064 as well bad, i think it should be divideable by 16

same for 808x792

[indigo carbon]

yep, this was the issue 😄

[clever verge]

Had to try it.

[soft zealot]

pass

[native charm]

Starry sky, sailboat, set sail

[steady grove]

arch imo . rolling release means its easier to update to newest available libraries when needed

[uncut fiber]

Like from some animated movie.

Nice alarm clock 🙂

[steady grove]

small controlnets https://www.reddit.com/r/StableDiffusion/comments/15ss1hv/57x_smaller_sdxl_controlnet_models/

[quartz sequoia]

@visual glade is there an auto node connect setting or mode? similar to how blender would work? just drag the node on the noodle and it connects the rest

[steady grove]

i'd love to drag nodes into the middle of a wire/noodle

i keep expecting it too. like i'll click a wire and select add reroute, thinking it'll put the reroute into the middle of that wire. nope

[quartz sequoia]

its very not LoRa friendly..

cuz if i have complex workflow

... its gonna be hell.. always connect and reconnecting the model and clip to everything that requires those

and if i want to staaaaaack

[steady grove]

i saw ctrl+b mentioned the other night and have started using it for loras

acts like the node isn't there then

[rustic garnet]

you can use reroute nodes for that

[soft zealot]

ComfyRoll has a LOA Loader with an off switch

[uncut fiber]

High native res. is best

[quartz sequoia]

true but then i gotta rerout the rerout

[quartz sequoia]

its good... but still the noodle for that last lora is gonna take a bit to connect to any node that require both the model and node

[steady grove]

even on XL i find better results come from primary low resolution pass. Gives the second some information to chew on. As usual, there's lots of options and barely a month into release means we don't know best practices yet

[uncut fiber]

yes native resolution should give best results. Check what native res. can do without upscaling.
#✨｜sdxl message
but yes as well getting weird results, just it take say 30 second, so why not 🙂
And height is native 1024

[uncut fiber]

weird it is not trained on 1024x576 because it is 1,7777777777 ratio, at least not seeing it there

[empty orbit]

do textual inversions work betwen sd1.5 and sdxl?

[steady grove]

nope

[empty orbit]

any good negative ones yet for sdxl

[steady grove]

theres a few but their effects are questionable

[empty orbit]

alright lol

[pure hazel]

MJ is sooo effing biased. It does what ever it wasnts 99% of the time. You have almost no control. It's just an 'easy' button with no creativity.

[uncut fiber]

should be evil fairy 🙂 doesnt look so evil

[steady grove]

i wouldn't trust her

[cinder aspen]

Its all many people need tho

Artists here and there

Some do use MJ with Firefly tho or some SD web UI

[steady grove]

for professional artists i dont think its a good tool. they take a lot of your intellectual property rights away in their terms of service. it's a huge trade off. https://docs.midjourney.com/docs/terms-of-service

for a toy, for naive people, $30 a month seems fair

[cinder aspen]

For artists Firefly is by far the best bet

[steady grove]

adobe will dominate that's true. Professionals would be idiots to rely on MJ. Especially with their royalty clauses

[cinder aspen]

But some use MJ atm as the image generator and then edit with generative fill

For me that is not a option anymore anyway

[steady grove]

Yeah, those people are idiots. MJ owns that image and allows them to use it

Why would a professional willfully enter into that contract?

I got a bridge to sell. Anyone wanna buy london bridge? once in a lifetime offer

[teal stratus]

Is there a pay option for the discord thing or is it just freeish?

[cinder aspen]

They dont care about the generated image being used on canvas because it very unlikely will be

[steady grove]

Care or not. MJ owns the IP and allows the users to use it. That's a huge legal liability that you're agreeing to

[cinder aspen]

But for majority of artists Firefly definitelly the one to go for several reasons

[steady grove]

"but they don't flex that ownership" is a level of trust that i can't understand

[cinder aspen]

As soon as its out of beta the game is over for the niche target

[uncut fiber]

somebody posted how many generated images, SD was far most by big margin. Others 3 had aroun 1 billion. Unfortunately cant find it, SD i think like 100+Billions

[steady grove]

SD is freely available and runs on consumer grade hardware so it's an obvious data result

[uncut fiber]

Yes it is.

[cinder aspen]

Yeah but thats for the generated image

[steady grove]

and all derivative works

[teal stratus]

I don't see how adobe is better

[cinder aspen]

When you take the image and use it as idea or reference MJ has nothing to do with it anymore

[steady grove]

licensing terms. ux. integration. support

[steady grove]

why pay MJ for that? just get a free result from google if all you need is inspiration or a reference.

$30 a month is steep for inspiration

[teal stratus]

It can't even make a woman om fire because that's against the terms.

[uncut fiber]

and somebody said, it can make ugly people

[cinder aspen]

Thats what most artists do, Pinterest, Artstation, Behance are still leading in that way. And now with Firefly it goes beyond that

[steady grove]

you've decided it's not a good system for you. that's okay. These controls won't matter to most using it. you'll see. Adobe will craft them better over time too

[steady grove]

exactly my point. MJ for those purposes is just throwing money away

[cinder aspen]

I pay already easy over 200€/month for art software, paying 30-60€ just for generating a image without any control anf customization is not a option for me anymore

[steady grove]

suit yourself

[teal stratus]

Adobe is a brand that I'm sure not many people trust. I don't think that because adobe has AI now it'll do anything. There will be just 1 button in Photoshop that lets you enter text and then it does AI stuff.

[steady grove]

you don't think people trust adobe?

maybe out of touch

[cinder aspen]

Adobe is trusted very much even when hated, or at least tolerated

[teal stratus]

I don't. Look at penpot and figma. Adobe bought figma and people were like fuck this shit.

[steady grove]

whats figma

[teal stratus]

What's adobe

🤔

[cinder aspen]

Remember when Adobe went to subscription model, it was said they are about to witness exodus and die

You know the epilogue?

[teal stratus]

What's figma. you serious bro.

[cinder aspen]

More profit than ever

Barely anyone left Adobe

[steady grove]

lol

[teal stratus]

I use adobe and don't pay for adobe. That subscription model just went in one ear and out the other.

[steady grove]

point at any design studio doing professional work. 99% chance they had adobe licensing cost in their business plan that they proposed to the bank when they were beginning.

[teal stratus]

Also normies don't think they have a choice.

[steady grove]

don't admit to pirating adobe in public forums. their legal teams are ruthless

if they sniff someone making money with pirated osftware, you'll get a very heavy letter

[teal stratus]

Too bad.

[steady grove]

it's not a problem for me. it's just good advice. don't be dumb hah

[teal stratus]

Heavy letters are good for heating up a house.

[steady grove]

don't admit to crimes

[teal stratus]

I logged off 500 prostitutes in Belgium 700 years ago. Is that also a crime?

[steady grove]

no not really. burning paper is a big nothing burger for that goal. Source: I use wood to heat my house. Paper just doesn't cut it.

[teal stratus]

A large company wanting to sue me I don't care about.

[steady grove]

look, i don't care how hard you think you are. cut the act.

flexing like that just makes me think you're soft

[cinder aspen]

Normies have a choice im a normie and i pay for both Creative Cloud and Substance and later eventually Adobe Stock

[teal stratus]

Not acting dude. 700 years ago that was me man. Just over here listing all the heinous crimes I've committed over the last 700 years.

[cinder aspen]

Its affordable

[steady grove]

delusional people sure can spiral

[teal stratus]

I prefer to not pay for it. I'd prefer to see adobe no longer make money because it's a terrible company.

[steady grove]

meanwhile, still pays for nestle products

[cinder aspen]

Someone else might argue Stability AI is terrible, yet there are people glorifying them

[teal stratus]

I don't eat goy slop

[steady grove]

"they're a terrible company" moral outrage is rarely held with vindication

[teal stratus]

Stability AI isn't that bad.

[hardy cipher]

what are you baby boys fighting about?

[cinder aspen]

Or people that hate the subscription model for softwares, yet pay on a daily basis for cigarettes, alcohol, Netflix, etc

[steady grove]

apparantly this guy is so hard that he pirates adobe and kills prostitutes in former lives

[teal stratus]

He's butt frustrated that adobe is shit and that he pays for something he shouldn't giving a company that makes billions more billions.

[steady grove]

we were just discussing the merrits of different services, and this guy is like "i'm hard af"

[cinder aspen]

What artists would use, MJ, Adobe Firefly or SD and then controversies around Adobe

[teal stratus]

At least you got it right.

[cinder aspen]

Why shouldnt i pay Adobe if i profit from them myself and get what i need

[steady grove]

do note that me recognizing your delusional claims is not a validation of those claims

[hardy cipher]

I mean, I pay for adobe. but only because I kept pirating it and eventually they gave me a popup that gave me adobe suite for 30 dollars a month "please, just give us some money, guy"

[teal stratus]

Exactly. You glad that new switch game came out. I think I can do a AI thing with bald head and switch game and unkempt beard. Frozen normie cuck look in eyes.

[hardy cipher]

so I figured that was fair

[steady grove]

i wonder where all that specificity came from

[teal stratus]

I don't let my pirate computer touch the internet.

[hardy cipher]

oh, okay

so you're from 1985

[cinder aspen]

I got mine for 35€/month for the first year and now that the discount is over they still make me pay less than usual

[steady grove]

no he's being sarcastic. If i had to guess, 2005.

[cinder aspen]

I didnt even have to make a deal with them now, they did it

[teal stratus]

I'm no zoomer

[teal stratus]

I'd just prefer to give them nothing. I still have the newest software. Paid nothing.

[hardy cipher]

my philosophy is, if oyu want to pay for something then pay for it. if you don't, then don't. and I don't care either way as it has no bearing or impact on my life whatsoever

[steady grove]

generational labels mean a lot less as soon as you're outside of highschool

[teal stratus]

This guy. Best guy.

[teal stratus]

You need to learn what different generations of people are and mean.

[cinder aspen]

Thats your decision, i prefer to pay and also get support etc on that. I will earn secondary money to make the costs disappear anyway in best case scenario

[steady grove]

true facts. the conversation was less about that though. it started as this guy saying nobody trusts adobe anymore. but then when tested, he went down a tangent about killing hundreds of prostitutes

[steady grove]

edgy troll is edgy

[hardy cipher]

adobe did that or he did?

[teal stratus]

To me it's like BMW saying ok let's charge extra per month for heated seats. The subscription model and not sticking with the you bought it and it's yours is not great.

[hardy cipher]

are you sure this isn't john mcaffee's ghost?

[cinder aspen]

Its a great deal for many tho

Who dont pay all at once

[teal stratus]

That dude went to South America and had a good time before he died.

[steady grove]

he claims he killed hundreds 700 years ago. so yeah, stop entertaining this one. he's edge to the max

[hardy cipher]

ooh, that's interesting

so he was a mayan, or aztec or something?

[cinder aspen]

At the end of the day use what you want

[teal stratus]

I guess whatever I write can be considered a crime and I can't write re tarded

[cinder aspen]

But why not boycott Adobe if hating them

Seems like that explains also why Adobe is still popular even through hate

[steady grove]

no. europe. super edge to the max. probably believes aryan race were white people.

[teal stratus]

Because affinity sucks balls.

[hardy cipher]

there are a lot of things in this world, and some of them I need to worry about. too many of them worry me. so I'm not going to waste my time worrying about some goof on the internet

[cinder aspen]

They arent bad at all tho

But with AI nobody beats Adobe in that sectpr

Sector

[hardy cipher]

I don't know. what's their latest AI triumphs? I haven't even messed with it for a while

[teal stratus]

Figma was bought by adobe and any prototype that's large enough to be small app sized slows figma down to shit let alone adove xd which sucks ass.

[hardy cipher]

I just know they could add emotion to someone's face and make it extremely offputting

[teal stratus]

Adobe firefly literally can't light a woman on fire.

[hardy cipher]

here, let me set this to 1 percent happy

[steady grove]

Adobe have been leaders in the computer vision field for decades

[hardy cipher]

then it creates a monster

[steady grove]

sample size of 1 might highlight confirmation biases

[hardy cipher]

adobe audition is underrated

that's all I know

you can reverse the polarity of only certain frequencies. I'm not saying that's why they're underrated. but it's a pretty interesting thing to do

[half ivy]

Yusssss

https://www.reddit.com/r/comfyui/comments/15t3jzt/comfy_is_the_best_of_an_imperfect_set_of_uis/?utm_source=share&utm_medium=ios_app&utm_name=ioscss&utm_content=1&utm_term=1

Can we get a Lobe theme too

[vale eagle]

LUL

[hardy cipher]

I'm really over the UI war nonsense. people act like we all have a vested interest in it. that what you choose somehow puts you in that tribe. they think if you do prefer a particular UI they can angrily demand you explain it to them. I realize this is probably just a vocal minority of individuals with low emotional intelligence, but it gets old

[half ivy]

Naa Naa bruv

Comfy is life

[half ivy]

Back to more important things though frfr

[wet nacelle]

https://tenor.com/view/jerma-gif-20630865

[hardy cipher]

that's why your mom told you not to make those weird faces

might get stuck that way

[wet nacelle]

https://tenor.com/view/jerma-discord-this-thing-what-gif-21031620

[teal stratus]

Is there an AI that won't restrict you from getting anything out of it? Any AI I've tried all seems incredibly woke.

Adobe nonsense super great AI can't make an image of a woman on fire because that's illegal to them. Most of the other ones including Stable Diffusion I guess Rick Roll you if your prompt results in something it doesn't agree with. I get that I can self host Stable Diffusion and I'm down with that but is it still going to tard out when it comes up with something it doesn't like?

[spring fulcrum]

Can LoRAs be used in ComfyUI by including them in the prompt? Or will they only work if there is a LoRA node?

[wet nacelle]

node from what I understand and use

[steady grove]

you're hung up on this. It's dumb.

in the future, whenever you don't like something because you feel like it's woke, pause, and realize you're being dumb.

[teal stratus]

There will be a non woke AI from a non woke as shit company.

[crisp owl]

There are uncensored LLM's you can install locally which will give you uncensored text responses, and a locally installed SD will let you generate whatever your heart so desires, as long as you can prompt for it, and there is generally stuff trained on the subject.

[teal stratus]

How isn't it woke..Most if not all of them are woke. Look up Taye and how Microsoft had to end it because she was extremely racist.

[teal stratus]

We talking images too?

[crisp owl]

SD, yes

[soft zealot]

aye, the UI doesnt prevent you from prompting for a "naked woman covered in cocolate baing sprayed with whipped cream" the only limitation is on how much the model knows about women , choclolate & whipped cream

[teal stratus]

Right on man. Good times.

[soft zealot]

unlike MJ where thaat prompt would likely get censo0red before htting the model

[steady grove]

i'm not going to argue if something is woke or not. What the word means to me in the context you're using it is "I hate human rights"

[crisp owl]

Anything webhosted, or hosted by a company, will absolutely, 100%, include restrictions. It's a way for them to stay legally out of the mud.

[teal stratus]

It's not the UI, it's whatever they have running as the stuff underneath that. That's obvious. It's just lame when you prompt and it just ignores it and makes something else. If anything you're charging people per generation of images and if you don't tell people that it won't be included then that's really fucked.

[teal stratus]

Number 4 looks like she has the downs.

[soft zealot]

if you go into any paid for service then unless it says otherwise I would always presume it to be prudish so as to avoid dissapointment

[vital ermine]

Not judging, but has something

[teal stratus]

I just don't care about any of that. If anything I'd do that. Unfiltered AI because it doesn't exist. Do you have any hints as to where I could check something out so I can find what I'm after?

[crisp owl]

Example of a webhosted place changing prompts.
blacked out because the linked post has an nsfw-ish image
||https://www.reddit.com/r/StableDiffusion/comments/14e4kdb/i_run_a_free_stable_diffusion_bot_i_have_fun/||

[teal stratus]

It's white.

[wet nacelle]

I'm white

https://tenor.com/view/asombrado-gif-26063609

[crisp owl]

I haven't looked at llm's in a long while, I downlaoded one earlier this year which wasn't censored

[crisp owl]

click it

[teal stratus]

I did. I'll read it. Reading now.

Thanks

[vital ermine]

That seems a bit hit or miss to me

[hardy cipher]

that's obviously the best one

it's the most letters and the highest number

[vital ermine]

Same everything only back to my unipc

[hardy cipher]

kanye surely owns that

[wet nacelle]

Ye when he waits for Giger to die just so that he can take his artistic likeness.

[hardy cipher]

giger is cool and all, but not a big fan of including his name in renders. seems to take them over

[hardy cipher]

alpha giger

[uncut fiber]

could be fun merge this author with Giger. This author i mean all knows.

[visual glade]

Should be fixed now

keeping stuff in vram shouldn't take much power at all

[cedar lark]

TY, confirmed it's working again with the advanced module with negative conditioning

[uncut fiber]

@visual glade i am deleting pictures directly from folder, and it is terible mess then when after generation it shows deleted image which is kept in memory 🙂 So i have to look in folder again.
No problem with it, just not sure how much my 8GB can hold in memory to not affect performance.
Thank you for reading this, and sorry to tagged you.

[high skiff]

It's weird, when I have The files for SDXL saved in my VRAM, my GPU goes up to boost clocks

[cinder aspen]

Woman on fire is censored

[molten gull]

hey @vital ermine 🙂 it#s been a while 🙂 all good?

[vital ermine]

all good.

[teal stratus]

That's pretty cool

[teal stratus]

That's dumb as shit. I don't want to make stuff for children that are 8 years old.

Is there a paid version for discord? I thought it wasn't free to generate images yet it seems to be free?

[hardy cipher]

are you guys trying to create edgelord images using the discord bot?

lol

lots of options there

[cinder aspen]

Some of thingd will be unbanned, as a user you can report that. Also, try to formulate it differently

But dont expect Adobe to allow gore in Firefly

But later they will make you able to train on your own images

[hardy cipher]

oh, adobe

well of course

that's just silly

just look at their terms of service. if you're going to be sending requests through their servers there will obviously be rules in place

[teal stratus]

Adobe can't even do that.

[hardy cipher]

so don't use it

[upbeat summit]

I created a script that when I'm done with adobe work, I run it to kill all the adobe services and tasks. it's crazy what they keep running in the background.

[hardy cipher]

oh yeah, you just have to look at your background processes

[teal stratus]

I had to tell it in the prompt to cover her in fire. Not that she was on fire.

[hardy cipher]

that's cool though. I need a script like that. I also like how it'll just leave 100 gb of temp files on my drive

that I then manually delete

[teal stratus]

Right and why do they keep it running? 🤔

[hardy cipher]

same with google drive. "here, you want 100 gb of trash data right?"

[upbeat summit]

mostly to have faster program start-ups, don't care to shut it down, license checks and collect system info 😉

[hardy cipher]

also "pip cache purge" on a regular basis. I'm not sure why all these software companies and creators think I want to trade significant amounts of storage space in order to save a few seconds of time every so often. the tradeoff really isn't reasonable

[teal stratus]

She's a witch burn her! Oh well she's already on fire.

[hardy cipher]

and I don't even want those things cached

[teal stratus]

Probably because they sell everything you do. Oh he likes coffee and red heads.

[hardy cipher]

probably

[vital ermine]

anyone know of a program that will take an image and slice it up for me into squares I tell it what size?

[teal stratus]

https://imagy.app/split-image/

[vital ermine]

thank you

[teal stratus]

Shit works

[teal stratus]

No problem

Why do you want to split up the image into different squares?

[azure oxide]

tokyo jab on reddit's method of animating uses that technique

neat site

[teal stratus]

I think I already know.

[cinder aspen]

Maybe try another prompt for same result

[teal stratus]

She looks mad she doesn't have cheeseburger shoes.

[vital ermine]

a technique to train an actual style when you have a limited number of images to work with

[teal stratus]

Good enough.

I thought you were making a puzzle.

[vital ermine]

no, but that would be cool

[strange mist]

this is def best in the series 💯

[thorny frost]

hi! a question: is any way to pause the queue in ComyUI and resume it at will. That would be great for workflows with uopscalers and such. You could see the preview of the lowres image and then decide if continue with to the upscale model or discard it and try another generation

[woeful patio]

does anyone know of a good way to colorize black and white manga?

[empty orbit]

controlnet probably

[empty orbit]

i love this lol

what prompt did you use

[hardy cipher]

mack got the best prompts, lol

[woeful patio]

anyone have a link to where i can download the 1.5 controlnet models for a1111?

[hardy cipher]

should be on huggingsface somewhere

[empty orbit]

https://github.com/Mikubill/sd-webui-controlnet

[hardy cipher]

I don't think they're UI specific

and github

[steady grove]

the controlnet extension ahs no sdxl support

[empty orbit]

install instructions should link anything else you need

[steady grove]

hardcoded to disable when sdxl is detected

[empty orbit]

pretty sure they're going to release it soon

but i think its still in progress

[hardy cipher]

I don't know about all that

I used a 1.5 canny controlnet model with sdxl. and it worked, kinda, lol

[empty orbit]

no looks like its done i found this https://huggingface.co/diffusers/controlnet-canny-sdxl-1.0

[hardy cipher]

not in a way that did anything good

[steady grove]

also, comfyui's controlnet is broken for me today. i updated now it's not working.

[hardy cipher]

but it definitely did something

[steady grove]

that error comes up whenever i have controlnet in the mix

[wet nacelle]

Are you using windows_portable?

[empty orbit]

if you install comfyui does it install the whole stablediffusion stuff again or can i just point it to an install

[hardy cipher]

no. although I have it installed. just stopped going through the portable startup

[steady grove]

no i found the actual line of code that detects sdxl and disables it. i'm not sure what you did but the extension is coded to know it's sdxl and just not turn on

theres even a message in the terminal about it

which is what i used to find the line

[jovial pasture]

No one use A1111 collab?

[hardy cipher]

I'd imagine it's new to the model then? I have models from probably june or something

[hardy cipher]

thelastben. I used to use it. it'd break constantly

[steady grove]

controlnet extension isn't a model it's code

[hardy cipher]

what controlnet extension? I'm guessing a1111

well I'm not using that

[jovial pasture]

This one! the best https://colab.research.google.com/github/sagiodev/stablediffusion_webui/blob/master/StableDiffusionUI_ngrok_sagiodev.ipynb#scrollTo=gJBtadxGOK7b

[steady grove]

you must've had a 1.5 model loaded and didn't realize it. controlnet 1.5 models just mathematically are not compatible with the new model

[wet nacelle]

I found that many things are missing from the old install of Comfy. I may be out of date on my side though.

When I check for updates through the manager it gives and error saying that it can not find the git folder.

When I installed portable it worked fine.

[jovial pasture]

controlnet full 1.5 already

[hardy cipher]

I manually updated the manager

git pulled

[wet nacelle]

ah

[hardy cipher]

no, I didn't. I realize you are sure I'm wrong

are you using a1111?

[steady grove]

i think whatever comfy changed last night to make controlnet work on lower vram cards, broke it for me. i'm confused by this error and don't know where it's happening module
module \'comfy.model_management\' has no attribute \'load_if_low_vram\'

[hardy cipher]

what's the rest of the error code?

[steady grove]

weird flex but ok. i dont want to argue but you gotta know, 1.5 controlnets won't work with sdxl. you were mistaken is the best case scenario.

[hardy cipher]

my friend, I'm not saying it worked correctly. I'm saying it was clear that it was impacting the output

[steady grove]

it wouldn't. the architectures are entirely different.

[wet nacelle]

"ALRIGHT YOU TWO! CALLM! THA FUHK DUHOWN!"

[steady grove]

why always people here hyping anger tones? Now the well is poisoned. I don't think a conversation is going to happen

[hardy cipher]

I just don't know why someone is so adamant about telling me I didn't do something, lol

I mean, I'm not angry, I just don't like being essentially called a liar

[steady grove]

helping you understand is alll. it's not an argument. if i had to guess you used img2img and saw vauge closeness. The extension is literally hard coded to not turn on. if it did it would just generate tensor errors

[hardy cipher]

I'm not dumb

[steady grove]

but now we're here and acting like i'm being hostile. it's just.. sigh

[hardy cipher]

well it's insulting

[visual glade]

I assume you are using a custom node?

[steady grove]

don't be so defensive. being mistaken is just an opportunity to be right

[steady grove]

maybe. they're not activated in my workflow far as i can tell. thanks for looking

[wet nacelle]

You just typed out the word "sigh"

That's not a normal thing to do as a human.

[visual glade]

I did a search and that function shouldn't be called from anywhere in the latest code

[hardy cipher]

ever hear someone say "L O L" outloud?

here, let me show you, iceycold

and you tell me

what it's doing

[steady grove]

i'll reinstall. but i'll come back later because i'm being dogpiled suddenly

[wet nacelle]

Used to often. It's now just the letters mashed to make a word.

[steady grove]

this is dumb

[hardy cipher]

really, I've got it loaded up

[azure oxide]

is there a setting to save as jpg in comfy? i read that stableswarm does it so there's gotta be a way right

[hardy cipher]

1.5 canny model

sdxl checkpoint

[steady grove]

dont bother i lost all respect for you when you recruited mackelmore and others to hype up your anger

[hardy cipher]

huh?

[steady grove]

forget it i said

[wet nacelle]

?

[hardy cipher]

recruit?

[wet nacelle]

McAndrews much?

[steady grove]

" ?? whatever do you mean ?? what could you possibly be implying ?? " yeah. forget it

[hardy cipher]

iceycold, you realize I was just going to show you that I'm not wrong

after you went out of your way to tell me I was wrong

[wet nacelle]

dis nword(slang) right here.

[steady grove]

@visual glade you're staff. this is outta control

[clever verge]

@visual glade Thanks for adding --disable-smart-memory, old way works better when upscaling 4 batch 4x upscale!

[wet nacelle]

Don't at them

[steady grove]

i can't even focus on my issues cause of this harassment

[hardy cipher]

now you see, as I said, it's not pretty, but it very clearly is responding to the controlnet data

exactly what I said

is this guy trolling or what?

[clever verge]

Would be even better if you could switch memory management mode from the GUI.

[wet nacelle]

I am lost and confused.

I just wanted to interact with the drama.

I saw you type out the word "sigh" in a literal way and wanted to insult you because of it.

What I know is that some 1.5 stuff has been able to work with XL by force.

Ya know what I'm saying? @steady grove

[hardy cipher]

strangly toxic, random behavior

tbh I kind of wonder what sort of weirdness I could come up with using that setup

[steady grove]

That is not at all working and thats very clearly denoised errors. this is dumb.

[hardy cipher]

k

[visual glade]

I'm staff but I'm not admin and I'm not really following this discussion, you need to report it

[hardy cipher]

it's very bizarre behavior

[visual glade]

by works better do you mean it OOMs if you don't do that?

[steady grove]

theres no report anywhere i thought we're just expected to ping people

its what i saw a mod say once before

[wet nacelle]

I just wanted to make a small insult at you.

[hardy cipher]

I was just commenting on the weird results I got. and I get told I'm wrong and that it's hardcorded into the extension? lol. what extension exactly?

[steady grove]

looks like you're allowed to get away with that around here big gunner. toxicity thrives. holy shit i try to explain that controlnet doesn't work that way and it goes to this place

a lot of respect lost today

[hardy cipher]

so that's just random too?

did I jailbreak something?

[wet nacelle]

I called you "not human"

I see this as a nonsensical way of demeaning another person. For a person to call it toxic is just odd to me.

[clever verge]

Yeah, I can keep the important stuff in VRAM using the older way. New way moves active stuff to RAM and it takes forever to complete as expected.

[steady grove]

you misunderstand is all. insisting you're right when you misunderstand is a problem i'm not about to engage

[hardy cipher]

you seem to have some emotional intelligence deficiencies so I'll go ahead and leave you alone. please have a blessed day

[azure oxide]

@visual glade any insight on this? i feel like im missing an obvious command line argument lol

[wet nacelle]

You are still lighting a fire man.

[visual glade]

comfyui doesn't support natively saving in jpg because I would have to implement loading the workflow in the UI from a jpg

[azure oxide]

thats what i was hoping you wouldnt say

[hardy cipher]

lol, well it was really bizarre to have someone tell me I was wrong about such a thing. it caught me off guard

[steady grove]

your dogwhistle needs tuning. i can still hear it

[hardy cipher]

I expected some sort of discourse and resolution, as is normal in such situations

[wet nacelle]

I know what you mean. What I do instead of drawing it out is just call the other person stupid and dumb and ignore them from there.

[azure oxide]

man are you two really continuing even though you both said youd drop it lmao 🙄

[hardy cipher]

https://tenor.com/view/ugh-annoyed-face-gif-18341274

[azure oxide]

it was a simple communication error

[visual glade]

what's your GPU?

[wet nacelle]

I then make myself look worse.

[steady grove]

casual racism is a new topic to discuss. one the mods don't take seriously enough

[azure oxide]

no need for this to be so blown and dragged out

[hardy cipher]

I agree, and I didn't mean to clutter things up with nonsense. that's my bad

[wet nacelle]

For the record I have from this point taken myself out from this discourse and will no longer be interacting with this.

[soft zealot]

One advantage of a slower card. Don’t like what you see in the preview window you have time to cancel that generation and let the next one load lol

[clever verge]

RTX 4090, torch-2.1.0.dev20230814+cu121

[visual glade]

so it's spilling to regular ram?

[clever verge]

Correct

[visual glade]

are you on linux or windows?

[gloomy barn]

https://www.youtube.com/watch?v=5xs-5DaewAA

[clever verge]

Windows 11

[hardy cipher]

my friend's little brother got super into those text based MUD games back in the day, and he'd do things like say "smile" instead of actually smile

[visual glade]

so you are doing a non latent hiresfix workflow with 4x upscale?

[gloomy barn]

That's funny.

[clever verge]

I'm basically doing what Sytans upscale workflow does

[gloomy barn]

/shakes my head

[hardy cipher]

nod

[clever verge]

Pretty basic

Might have been using a LoRA though, I have to test it again to refresh my memory. But I'm doing training now so have to wait.

[visual glade]

can you test it now to see if it's fixed?

[clever verge]

I'll test when batch run is done, I have 1000 steps left.

[woeful patio]

is lineart controlnet out for sdxl?

[vale eagle]

soon

[vocal stream]

is 32GB ram basically untenable for using stuff like refiner and upscaler with a 4090? Really dont want to buy more ddr4 ram to just use for like 6 months but no workflows except the super basic one seem to work with 32gb

[vital ermine]

depends. In Auto1111 even my 48GB is iffy

in comfyui 32 is enough

[vocal stream]

in auto111 I constantly OOM with 32gb but it sometimes work if the settings are right from load. In comfy base works but serge and more complex workflows OOM, too

or at least any workflow that uses stuff like refiner

[vital ermine]

hI use dev branch as the main branch OOM me all the time and I had 128GB page file. I limited it to 16GB and received a fat BSOD

it is better in the dev, but I watched it got dangerously close to 40GB

[vocal stream]

I do use dev branch for auto but it's stil hit and miss and I get the OOM message especially if I swap models

[vital ermine]

tbh, I don't know of anyone who still uses Auto1111 that uses the refiner. Just too risky

[vocal stream]

yeh for auto i dont even need the refiner to OOM, just loading one of the models from civtai that say they dont require it

[vital ermine]

besides, the way Auto does it is load, unload, load, unload then load base back. That is 100% shit

[vocal stream]

in auto with 32 I legit have to stop it, change it from the json to a different model and then start it since I can never load a different model after it loads on start whatever i had last

[vital ermine]

oh, 4090 here and 48gb and I don't use models besides base. I train my own loras if I need them. Not tried DB yet

[vocal stream]

super clunky

[vital ermine]

if Auto just loaded both models, as comfy does, a lot of issues would vanish

[hallow summit]

auto can hold more models, too - it's available in options (at least on dev)

[vital ermine]

sure but he doesn't

refiner base he still hasn't done it

on my 1060 I held in mem a lot

[hallow summit]

it does, check dev branch

[vital ermine]

I am on the dev brach and he loads, unloads then loads back up. That is crap

[vocal stream]

hm seems like my pagefile is 16gb, maybe I need to up that to not OOM as much (tho seems iffy to be going to disc instead of RAM)

[hallow summit]

that might be caused by low VRAM

for me it holds both base and refiner

[vital ermine]

Ahem 4090 24GB and 48GB system ram

[hallow summit]

set COMMANDLINE_ARGS=--opt-sub-quad-attention --xformers --medvram --no-half-vae
set ATTN_PRECISION=fp16

(32GB RAM, 16GB VRAM)

stays under 8GB RAM VRAM usage when generating 1024x1024 using SDXL base with refiner

[vital ermine]

well, tbh Auto sucks ass in his memory management and Comfy blows his ass out of the water in that department. I don't use Comfy because he doesn't allow my iA3 loras else I would go back. I am really underwhelmed by Auto's dev branch.

[vocal stream]

they do seem to be working a lot (given that everyday when I pull there's new commits) that I've ben expecting for a while to catchup but if anything now more stuff are broken on dev

(like deforum, roop etc)

[vital ermine]

well, base refiner I am using up to 36GB system ram

[hallow summit]

yeah, ComfyUI does it pretty well (though I prefer to launch it without smart memory management - it stays under 8GB then, too)

[vital ermine]

I haven't tried that

[vocal stream]

hm I havent messed with flags

with a 4090 I kind of thought id just be able to run things straight ngl

[vital ermine]

nope

things are still catching up to ADA btw

[steel mist]

This discord is bad for one’s personal finances

[vital ermine]

Even Nvidia acknowledges issues in their drivers for us so they said in a future release

[vital ermine]

I saved for 7 months and still 500 short. 😦

[steel mist]

I’m gonna need to check my ram use

[hallow summit]

for basic use cases (when I don't need to play with the workflow) I recently started using Fooocus (by lllyasviel) - it's largely based on ComfyUI code and has nice integrated sampler doing both base and refiner

[vocal stream]

tbf I cant really understand how RAM can be a limiter if it's not trying to be clever - with a 4090 can't you just .cuda() all the models and load things one by one and never have to OOM

[steel mist]

Problem is I’m running the PC headless and I never bother keeping the remote monitor up

[vital ermine]

Nerdy Rodent and now I can't get that out of my head.

[vocal stream]

like what is it even using so much ram for after it first loads the models, why are they both in vram and ram

[vital ermine]

good question

[steel mist]

I mean folks we are talking about python here right

Py is great but not known for it’s speed or memory efficiency?

[vital ermine]

Sadly, yes

[hallow summit]

(just had to tweak it a bit myself, as vanilla version doesn't allow to change even basic settings)

[vocal stream]

eh, for cuda stuff like this it's all C bindings (from pytorch in this case). Python adds minimal overhead

[steel mist]

My assumption (and I have not looked deep into it) was that python was still likely doing a lot of memory management

[hallow summit]

python is used for logic and UI stuff, all the heavy computations are accelerated by GPU

[visual glade]

what I hate about python is that I can't free()

[steel mist]

yeah for sure, I mean if this were interpreted python we’d be talking days to render an image

[vocal stream]

you can gc.collect() and del tho?

[visual glade]

del is "delete this if you want"

it's not a free()

[vocal stream]

if you do del and then gc.collect() right after in 'pure' python it's pretty much free() but here we're dealing with bindings to C and libs so you might just del your reference but it keeps staying there in the C lib or whatever

oh huh increasing the pagefile more did work (or maybe it's just the restart and me using 5gb or whatever less RAM because of it)

[clever verge]

Seems to be like before. It happens during the vae decode. Forgot about that.

[steady grove]

have reinstalled comfyui and updated it. trimmed down my workflow so aboslutely no custom nodes were in it. i'm still getting this error

[visual glade]

how big are you upscaling to?

[wet nacelle]

@visual gladeAre you the one that codded it to highlight a node in purple when it failed or was it somebody that made a node set?

If it was you that implemented that then why is it that color?

[steady grove]

i downloaded comfy from github again, unzipped it, the unziped version didn't have canny, so i ran the update bat file. i added the controlnet stuff to sytan's workflow, and i get that error

[clever verge]

1536x640 and a 4x upscale and then a ksampler base for a few steps then to the decoder

[visual glade]

"F:\ComfyUI_windows_portable\ComfyUI\custom_nodes\AIT\AITemplate\AITemplate.py
That's the source of the error

so you should update it

or remove it

[steady grove]

i'm not using them because they make nothing work. i'll just delete that

[hallow summit]

usually VAE decode takes most VRAM for higher resolutions. you can try using (testing) tiled VAE decode - but it messes up colors a bit

[steady grove]

yup. thanks comfy. looks like just being in the custom node folder is enough for AIT to break controlnet

i deleted it, changed nothing, and it works

[visual glade]

I'm not the one who put that color but what's wrong with it?

[clever verge]

It's happening with 1024x1024 even right after base + refiner when I vae decode it for preview before it's up to the upscaler

[wet nacelle]

Nothing

[hardy cipher]

using the weird 1.5 controlnet canny results as the latent image for a proper render produces some interesting results

[azure oxide]

You can recolor nodes by right clicking on them and selecting it

[visual glade]

so you do a 4096x4096 batch size 4?

[clever verge]

correct

[indigo carbon]

nah, that node works perfectly, likely not the issue

only issue with it is if you use it on SDXL and latent size isn't dividable by 16, it won't diffuse correctly

[steady grove]

i deleted the folder , changed nothing, controlnet works

[hardy cipher]

what's up with the custom colors in groups? am I missing something or do they come out very strange?

[steady grove]

wasn't using the node. the only thing that changed is comfyui didn't initialize that folder

[indigo carbon]

did you use Cnet on it? it doesn't support Cnet

[steady grove]

controlnet extension now works

[clever verge]

Not sure why it's using 23 out of 24 for this though. I'll switch to the old memory management and see what it uses.

[steady grove]

it was breaking controlnet without it even being in the workflow

[visual glade]

yeah that node tries patching the comfyui code I think

[steady grove]

wreckless

[indigo carbon]

very odd. well, you can always disable it using the manager..

[steady grove]

i deleted it

[hardy cipher]

you know, I kept getting an error related to that node as well when trying to use controlnet with sdxl

[azure oxide]

unsure, havent tried coloring groups myself

[hardy cipher]

figured I just didn't know what I was doing and let it be

[steady grove]

you know what my totally useless feature request is? better node coloring. like, cool interpolation effects

also, you select a custom color, save a color in the palette, open a new custom color palette, saved custom color is gone

[azure oxide]

thats a 3rd party thing that someone could easily make up

[indigo carbon]

also batch count would be awesome

[hardy cipher]

super minor issue thatt I'm sure is easily resolvable. just something I noticed

when I organize things I get a bit ocd with the colors, lol

[steady grove]

i've done stuff with d3.js before so it might be something i'd mess with if i was sharper

[visual glade]

does it auto switch to tiled VAE decoding?

[hardy cipher]

I'm working on better understanding how the nodes work. even made my own quasi-useful node

[indigo carbon]

@visual glade will there be a batch count feature like A1111 had? (not batch size)

[visual glade]

that's in the "extra options"

[clever verge]

I seem to be able to reproduce even without the upscaler. Just base + refiner with euler and a LoRA. I'm at 8.6G with smart memory disabled right before the vae decode and don't see any increase after that. with smart mem enabled im at 13G right before the vae decode and when it's decoding it goes to 23 and get slow.

[azure oxide]

extra options also has the "generate infinitely" feature with that unlabeled box, psa

[indigo carbon]

I know, that's why I said "like A1111 had" it's different, when you do batch count like that it doesn't make a grid or let you see all gens from that batch, only the last one

[clever verge]

bypassed the lora (with ctrl+b) same issue, I'll try to rewire to remove it completely.

[soft zealot]

theres been a Batch Count like forever 🙂

[azure oxide]

caveat with his followup tho

[indigo carbon]

.

[hardy cipher]

hidden features

[soft zealot]

hardly call it hidden, tick a box and its right in front of you

[indigo carbon]

that's really not what I meant.. it doesn't show you all gens from the batch, only the last one. also no grid

[clever verge]

same issue with lora out of the flow

[crisp owl]

And you can click the other checkbox to basically generate forever

[soft zealot]

Its what was asked for , a batch countr, its there 🙂

What you meant to ask for as some sort of grid/gallery output

[visual glade]

can you show me the first few lines it prints when you start it?

[crisp owl]

oh rootbeer said that above, sorry didn't backread

[indigo carbon]

I don't know. I just thought it would be nice if I input batch count 4 I would get a grid with all 4 gens, not just the last..

[soft zealot]

Yes but you dint ask for that you asked for a Batch Count, which already exists 🙂

what you meant to ask for is a grid output like you get when using XY plot scripting

[indigo carbon]

that's not how batch count originally was. it didn't just make 4 images, it also showed in the output the 4 images that were requested

[clever verge]

from cold start?

[visual glade]

yes

the line with the GPU name and the ones beside it

[clever verge]

Total VRAM 24564 MB, total RAM 65456 MB
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce RTX 4090 : cudaMallocAsync
Using pytorch cross attention

Import times for custom nodes:
0.0 seconds: E:\Support\AI\ComfyUI\custom_nodes\sdxl_prompt_styler
0.0 seconds: E:\Support\AI\ComfyUI\custom_nodes\sdxl_primitive

[soft zealot]

I've no idea , Batch count has always for me generated X number of images sequentially (unless the Batch Size was greater than 1 in which case it would be X time Y images

[azure oxide]

yeah if it was bs5 for example you could click on 1 and cycle thru left and right on the preview in a1

[woeful patio]

can anyone help me with this error i get when launching a1111: stderr: ERROR: Ignored the following versions that require a different python version: 1.6.2 Requires-Python >=3.7,<3.10; 1.6.3 Requires-Python >=3.7,<3.10; 1.7.0 Requires-Python >=3.7,<3.10; 1.7.1 Requires-Python >=3.7,<3.10
ERROR: Could not find a version that satisfies the requirement tb-nightly (from gfpgan) (from versions: none)
ERROR: No matching distribution found for tb-nightly

[notice] A new release of pip available: 22.2.1 -> 23.2.1
[notice] To update, run: D:\sd2\stable-diffusion-webui\venv\Scripts\python.exe -m pip install --upgrade pip

[indigo carbon]

when A1111 was the way, batch count did exactly what I mentioned before. it's just disappointing that ComfyUI doesn't do that as well

[crisp owl]

To update, run: D:\sd2\stable-diffusion-webui\venv\Scripts\python.exe -m pip install --upgrade pip

[soft zealot]

ComfyUI is NOT any sort of prt or fork odf A11111 so you shouldnt expect things to be the same

you asked if there was a Btach Count, there is

[hardy cipher]

I like how a checkpoint model will load in about 5 seconds from my ssd, but takes about 5 minutes from my external hdd. bit of a difference there

[indigo carbon]

I didn't expect anything to be the same.. I just said I really miss that feature.

that's all

[soft zealot]

FYI if you increase Batch Size to Greater than 1 then you get multiple results

[azure oxide]

tb-nightly doesnt sound like a default a1 implementation tbh altho i could be wrong

[soft zealot]

and yes that is a semi naked woman again to whoever doesnt like them, get over it

[indigo carbon]

yeah, well, AIT doesn't do batch size yet- so.. y'know

[soft zealot]

dont Use AIT then :p

[indigo carbon]

speed>comfortability IMO

[hardy cipher]

I just don't know how it really works

[soft zealot]

well I guess that s mor eimprortant to haveinga ll 4 images on view.

Which I doint see as a huge issue, I just keep an exploder window open to one side whilst Im using comfy * I can see the thumbnails there

[indigo carbon]

yeah, that's similar to what I do. the speed that AIT makes 2 images is way faster than bs2 without AIT, so I guess I'll stay like this. or I'll just make a batch count node

[visual glade]

are you using any command line options?

[clever verge]

None

[steady grove]

empty latent image node

[azure oxide]

Using pytorch cross attention
does comfyui automatically add this for 4090s? i had to manually put it in my .bat with my 3090

[visual glade]

when there's no xformers and pytorch is 2.0 or up

[azure oxide]

ahh kk

[indigo carbon]

isn't vanilla Xformers enabled by default?

[visual glade]

if you have it installed

[spring fulcrum]

@visual glade How do you add more options to an existing node? I would like to have more LoRAs on this node. Is there something in the JSON I can change to add more to it?

[crisp owl]

edit the .py file for that node, or do a node merge

[soft zealot]

I see civitai is taking a coffee break

[crisp owl]

yah they announced #1037799583784370199 message

[steady grove]

ooo secret servers i dont have access to

[crisp owl]

just civit

[woeful patio]

Im remaining with this error now when launching a1111:
stderr: ERROR: Ignored the following versions that require a different python version: 1.6.2 Requires-Python >=3.7,<3.10; 1.6.3 Requires-Python >=3.7,<3.10; 1.7.0 Requires-Python >=3.7,<3.10; 1.7.1 Requires-Python >=3.7,<3.10
ERROR: Could not find a version that satisfies the requirement tb-nightly (from gfpgan) (from versions: none)
ERROR: No matching distribution found for tb-nightly

[steady grove]

says i'm not allowed. i've never even been there lol wow

[soft zealot]

would help if I had looked lol

[steady grove]

i was preemptively banned? must've pissed off an admin's alt or something. i dont even know bro

[hardy cipher]

anyone using these nodes? is the RescalClassifierFreeGuidanceTest essentially noise offset?

[crisp owl]

nah it's because they have a "welcome" channel you have to agree to terms before the other channels open up. You have to click their invite link from the website

[steady grove]

remind me to not sign into civit using my discord account ever

[steady grove]

link has a big lock on it for me

[spring fulcrum]

How do I find where that is? I am currently using Searge 3.4 and Its in the custom nodes folder.

[crisp owl]

I don't have Searge pack, but you can do a grep search for that node name and see which .py file contains it

[hardy cipher]

I don't think there's a simple formula, Scorp. I'd suggest looking at other nodes that stack loras as well as the searge node and try to figure out how to combine them

[crisp owl]

it's not "simple" no. Especially if one doesn't quite know what they're doing

node merging would be the more basic route

ymmv with those

[steady grove]

i seent a node that allows you to call loras from the text box like automatic

[spring fulcrum]

I found it... Its in ui.py

hopefully it works

[hardy cipher]

might take some trial and error

at least for me it did

[spring fulcrum]

I added more loras here but i know in the json I might need to add index for the extra slots on the node itself

[hardy cipher]

I think you'll have to give the unique identifiers

[crisp owl]

yup they need to be unique

may have to map elsewhere if they are called in other places in the .py

[hardy cipher]

yeah, searge's nodes are kind of spread out as far as the code is concerned

[steady grove]

https://github.com/badjeff/comfyui_lora_tag_loader i found it. this might be what i start using. i like the idea of adding or removing a lora with the prompt so much more. maybe its just what i'm used to

[crisp owl]

I've done some node customization which were a nightmare to do. Others are super easy

[hardy cipher]

I just used the sample node along with similar nodes for reference

I mean, not saying it was super simple for me. but I'm not the best with python

I've actually been wanting to create a lora stack node similar the the efficiency lora stack, but with normal model/clip output

I appreciate all the node work people are doing, but kind of a bummer when the only option is to then use their sampler

[vast narwhal]

The rescale, it does have effect. I can't see the tonemapnoise helping

[hardy cipher]

I'm guessing it does some sort of cfg scheduling or clamping?

the tonemapnoise

[vast narwhal]

I need to learn not to update things

[midnight shuttle]

What did you update? 99% of the time, updating fixes more problems than it solves. Not updating is boomer thinking.

[vast narwhal]

There's a pdf explaining, but SDXL is already fixed, i guess

[worthy orbit]

Any simple inpainting workflow for Comfy, or Upscalling that doesn't get my GPU on fire? (RX 6700 XT)? I found really cumbersome Comfy for inpainting or changing from generation to fix/upscale

[vast narwhal]

I always update everything, the comfy and nodes, yesterday everything was working now that. I love updating but i think "working" is better

[midnight shuttle]

Looks like something increased your VRAM usage. Are you attempting a different resolution?

[vast narwhal]

No. I just loaded my saved workflow

[polar epoch]

Hah cool xD Tested civitA.I's web gen, and went crazy on prompts, went mad good to be first try lol

[compact hatch]

SDXL is a different model so this might not apply, but it generally desaturated 1.5 models unless they were trained using minsnr with v-loss scaling. In that case it helped prevent oversaturation, overly bright/dark generations, and black/white box outputs. I don't think it's really worth using unless you're using a vpred model.

[visual glade]

--disable-smart-memory if you want the old memory management behavior

[vast narwhal]

Looks like isn't that. In another workflow i got this. This one was also working yesterday

[soft zealot]

whats the chang ein Memory Management please?

[visual glade]

it tries to keep some things in GPU memory to increase speed and decrease regular ram usage a bit

[soft zealot]

ah ok so effectivly swapping less. TBH with 64Gb system RAM I'll probably disable this and keep the 11Gb of VRAM for when its needed 🙂

[eternal fog]

That Smart Memory feature is probably what's making it no longer freeze when I use LoRAs then

[placid vessel]

You ever found that out? since im trying to find out too

[eternal fog]

That won't be img2img, they are just using an Anime Model, likely on SD 1.5 looking at that and are using text prompting.

[placid vessel]

If you take a look at his tiktok some of them are like exact replicas of the character that I feel cant be in just text but

[eternal fog]

It could be.
If they were using img2img at all, it would be at a very very high denoise, otherwise it would just look like a mess.

[placid vessel]

[eternal fog]

It's not hard to do something similar to that

A roblox model isn't exactly complex so you just need to describe the clothes and colours

[eternal fog]

Can't believe you'd burn down a village like that

[vital ermine]

Prime real Estate

testing my lora

without the lora

[eternal fog]

What is your LoRA? Burning down villages LoRA.

[vital ermine]

I think I need a few more images so need a few more burning villages to really sew it up.

[steady grove]

seems a little tasteless

[calm bramble]

what settings is the discord bot using for sdxl?

[steady grove]

such an enigma whatever could it mean

[primal vault]

Don't change the existing node, you can make a copy of it and register it with a new name, but if you change the node in code you'll be in a lot of pain when that conflicts with the original node from my extension

[hardy cipher]

lol

just pray you don't get the aiohttp errors

[cinder aspen]

@vital erminelike the style

[vital ermine]

Thank you

[cinder aspen]

the 2nd last one has a sign or watermark 😄

bottom right

[vital ermine]

yep, even with my neg

[cinder aspen]

according to some it even adds stuff lol

[vital ermine]

I wouldn't doubt it, lol

[cinder aspen]

you still on 1.5 or 2.1?

[vital ermine]

XL

[cinder aspen]

oh

[vital ermine]

I know 768

[cinder aspen]

didnt you say you will stick to the older ones for some reason? 😄

[vital ermine]

I did but then I got my 4090

[cinder aspen]

O_O 😄

[vital ermine]

1060 sucked with xl

[cinder aspen]

i need that one as well

[cinder aspen]

RTX 2070 is a frustrating experience as well for my expectations

[vital ermine]

4090 is not without issues still, and next year I need to upgrade the pc for it

[cinder aspen]

i deleted SDXL + ComfyUI then

[cinder aspen]

who knows if something new gen will come out next year

[vital ermine]

I know Zen5 is 30% faster than Zen4. Already been talking about Zen6

My old system can't keep up with the card so I stutter like mad when I gen a batch or train

train isn't as bad since most of that is in the card but gen is ugh

Next cpu will be a 16 core 32t, no doubt about it

[cinder aspen]

wonder how expensive those will be

[vital ermine]

499

for zen5 I dunno as the entire AM5 setup is hellishly expensive

[heady vale]

AMD seem to focus a lot on multicore performance for CPUs but they are lacking in single thread workloads. They have over come that in gaming with X3D large cache designs, but single thread is still lacking

[vital ermine]

yep

they don't care it would seem

[cinder aspen]

i also hope for the fix for the issues with above 64 GB of DDR5 RAM

[vital ermine]

issue with that too?

mobos been able to do 128 for a while now

[cinder aspen]

yes

but DDR5

not DDR4

stability issues with everything above 64 GB

[vital ermine]

I was going to do 128 and be done with it on the new board so I hope next year I do it is fixed

[cinder aspen]

and one of the shops i want to buy from doesnt build them for you that way atm

because of this

they set a warning

[vital ermine]

That sucks

Not like ddr5 hasn't been out for a good while

[cinder aspen]

if its not fixed i have to stick to build upon 2x 32 GB until its fixed so i upgrade then or maybe something new comes

[vital ermine]

I am still not liking DDR5 being so slow still

their timings are slower than equiv DDR4. a 6000 CL30 is slower than 3200 CL14

That cl needs to come down some or leave and jack the freq up by a lot

[cinder aspen]

dunno man, i hope something happens before we both upgrade/build 😄

[cinder aspen]

im really trying to not spend over 6.000€ but it will be hard

maybe impossible

[elfin flare]

Huh?

[vital ermine]

I am doing all of this a piece at a time

[elfin flare]

A 4090 with exactly 64gb of DDR5 and a good CPU would net you what

like 3000-4000?

[vital ermine]

this card hit hard as I have saved for 7 months and still have to find 500

[elfin flare]

What do you mean 6000 euros

What are you buying for a PC that costs that much

[cinder aspen]

@vital erminebut you are upgrading, i will literally build from scratch

[cinder aspen]

monitor and other things as well

[elfin flare]

Please dont tell me you're one of the people that would buy a 300hz monitor LOL

[cinder aspen]

monitor is 700-2000 depending on decision

[elfin flare]

anything above 144hz is completely pointless imo

[cinder aspen]

300hz not necessarily, but 4K and high color accuracy are important

[elfin flare]

Why 4K?

Why not a 1440P display?

Thenagain

your budget is 6,000 and under

I guess that's fair

[cinder aspen]

if i build around high end parts then why not 😄 But im not building just for gaming

the other major part is creative stuff

[elfin flare]

Well if your intention is a hybrid

[cinder aspen]

2D and 3D art. Modeling, animating, simulating etc.

[elfin flare]

I assume you're going high-end gaming GPU