✨｜sdxl

[half ivy]

Wedding cakes LG!!

[crisp owl]

dc marvel

[uncut fiber]

not wedding cakes, just normal those 🙂

[hardy cipher]

wtf is this?

[tall surge]

So I am a bit confused about controlnet and these openpose.zips I found on model sites 🤔
Where do I actually put these images and json files?

I cant seem to find that info online while looking at guides

[hardy cipher]

using what interface?

[mossy canopy]

models/controlnet

[uncut fiber]

faberge egg 😄

[hardy cipher]

well, doubt you'd put images in the controlnet folder

they're probably images with comfy workflows in them

[tall surge]

Im guessing I can just put these folders in that controlnet folder?

[hardy cipher]

yeah, put them there

that makes perfect sense

[tall surge]

Honestly I'm just asking lol

[uncut fiber]

@half ivy those are wedding. We got some story about how cat and dog cooked cake and put in it all they like most, fish, soap simply everything 🙂

[hardy cipher]

wat?

[tall surge]

I have no clue how to even work controlnet yet

[hardy cipher]

well, those aren't controlnet models, they're images and json files

[tall surge]

I thought I needed controlnet for openpose?

[hardy cipher]

if there are controlnet models amongst them, they'll be much larger, and they won't be image or json files

[half ivy]

Errrmmmeerrggeerrrrdddd

[half ivy]

Join the club

[vital ermine]

I dunno, but I love it.

[hardy cipher]

put together a bizarro workflow

[mossy canopy]

any a770 user here?

[hardy cipher]

you know, if lumi hadn't ignored my first question I could help him

but alas, ignored, lol

[half ivy]

And they say we can’t handle aliens

[half ivy]

Can I have this picture?

[uncut fiber]

what picture? It is story. Going to check

[half ivy]

This one

Last one jewelry is better though

[hardy cipher]

it's evolving

[uncut fiber]

OT: @half ivy

[hardy cipher]

lol. man, these are so weird.

I still don't understand how they differ from unclip though. better results, but seem like the same concept

[uncut fiber]

reminds me POV renderer

[hardy cipher]

advanced bryce render

[tepid surge]

wooosh ✨

when humans nice to eachother

[hardy cipher]

there's some legitimately off the wall people that pass through this channel

[ionic dragon]

found it
https://docs.google.com/document/d/1BxdWqfBJ3QPggHnBCBx3QIkUpHnBWShd_s3zEt2dTLM/

[half ivy]

cant get over this, creativity on my face bruhhhh

not sure the prompt bt for sd to recognise and shape so many colours and dimensions but retain a face is cray cray frfr

[strong field]

it is wild

[half ivy]

love it

was gonna "those eyes" but then

[half ivy]

taters gonna tate bt feel xl should be 1 dim less, 1024 making too many duplicates like when yr dims are too big

[strong field]

its hit and miss, but its getting better

base is suprisingly good

[half ivy]

off the record, is she single?

[strong field]

and if all you have to do i inpaint the eyes....

[strong field]

fr fr

[half ivy]

i needs inpaint with control net

fkml

i need 3 control nets

[strong field]

https://tenor.com/view/lunch-lady-farley-hair-shake-gif-25291072

[half ivy]

ref, lineart realistic, and the newest 1 yet to be released (hope)

[strong field]

that signature tho

hahahah

[half ivy]

i know thast guy, i dont trust him btw

[half ivy]

signature series

[half ivy]

few understand \

question though, whats with all the stencil drawings

and also

whats up with steven seagul

LOL i hate love them

[hardy cipher]

I like to make things of dubious quality

that one still hs the signature

[half ivy]

bro thats quality of hating something you love, typical hate love but boosted

[half ivy]

an open challenge
Dont need yr workflow, just want to see whats possible

[hardy cipher]

using this with clipvision in my negative prompt seems to have improved image quality

[uncut fiber]

how to get full details about image in ComfUI, when using random numbers?

[hardy cipher]

not sure what you mean by that?

[uncut fiber]

i made my cfg random from 4-12 and i think that random number will be regenerated

[hardy cipher]

just trying to make future machines

[half ivy]

not too many redies

[hardy cipher]

oh, two of those are nodes I made

[hardy cipher]

they're not listed in comfy manager atm

[half ivy]

bruh

[hardy cipher]

I can link you if you want. but you really don't need them

just a convenient 10 stack lora loader, and a node that will calculate any aspect ratio for any number of pixels in a square

[half ivy]

that what you sent me, not the lora part

tab open ready for tomoz

[hardy cipher]

perhaps. yeah. I'm working on more. I have some ideas that might be useful

[half ivy]

an open letter to user interface management, contact my guy at lobe extension, tell them TIA

[half ivy]

that sounds cool af btw, wish i had that swag, but i have moral support and thats summin at least, LG Team!!!

[hardy cipher]

https://tenor.com/view/aerobics-dream-team-gif-20921971

[pine portal]

boo

this one is scary

its like "oi bruv u want some?"

[hardy cipher]

very regal lads

[ionic dragon]

does the refiner consume more vram?

[vital ermine]

yes, in auto he unloads base then loads refiner then loads base back but that has issues.

[ionic dragon]

the 3rd pass is 5times slower

[vital ermine]

I gave up on ComfyUI and went back to Auto1111 dev build. At least with it I can use my iA3 loras.

[ionic dragon]

whats iA3?

[vital ermine]

Not sure what LyCORIS did to Comfy but it must have been something for him to ignore DyLoRA and iA3

another type of training. 900k, yes k, for XL and 287k for 2.1. Really good.

my hope is someone can make an extension for comfy so they can be used.

[ionic dragon]

oh ok

[steady grove]

maybe there isn't a personal vendetta and comfy is just one man who works on one thing at a time?

turns out software development isn't the real housewives of github

[vital ermine]

of course I get the suspicion it isn't comfy it is SAI themselves since even Joe ignored the request but w/e. Wanna bet icey.

[steady grove]

Discord bets.... where you can't see the person's face or shake their hand... no thanks

[vital ermine]

¯_(ツ)_/¯

[steady grove]

@hard fractal wanna weigh in on this conspiracy theory ?

[vital ermine]

Yes, please do as I want to know precisely why it is ignored. Might be a real programming cause too.

[steady grove]

https://github.com/tripplyons/sd-ia3/blob/main/infer.py looks easy enough to implement. it's probably just a time issue and also nobody else cares about this.

show the benefits of ia3 and somebody will have it implemented in a day.

[visual glade]

iA3 isn't even a lora

[vital ermine]

FFS, small, fast to train, works wonderfully well. They have been shown but ignored.

[steady grove]

maam, calm down

[visual glade]

it requires modifying the unet architecture not just the weights

that's why it's not implemented

[vital ermine]

How does it work so well in Auto yet will not work in ComfyUI?

[steady grove]

cool. so not anything to do with hating the developer of ia3. wow that's so obvious.

we didn't make a bet but i think an admission of "i was wrong" is in order

[hardy cipher]

can't do everything at once?

[visual glade]

because I try not modifying the unet architecture because it's going to break future optimizations that I'm implementing

for example IA3 won't work with AITemplate

[steady grove]

Great thing about FOSS is that somebody can always fork it and do it themselves too.

[vital ermine]

Apparently your limited vision skimmed over that last part.

[wet nacelle]

Hey I'm using one of your workflows right now.

I can output a couple of times but then I get an error wall.

Does this have to do with the memory not leaving and instead staying when it shouldn't?

[steady grove]

what was this about then? /shrug

[visual glade]

since it doesn't work with AITemplate it means that IA3 is going to be over 50% slower than using regular loras

[steady grove]

😮 woah a real engineering reason

instead of a conspiracy assumption 😮

[visual glade]

so you should not use IA3

[vital ermine]

Now, wouldn't that have been great if that would have been said months ago or weeks ago instead> Yes, no, but whatever.

[steady grove]

https://tenor.com/view/pikachu-shocked-face-stunned-pokemon-shocked-not-shocked-omg-gif-24112152

[ionic dragon]

These all resolutions are supported by SD?

[steady grove]

/shrug i asked and it was answered in a minute

[vital ermine]

I effing asked, and brought it up over and over again and nothing. Silence is not golden when it is asked and the devs go deaf, dumb, and mute.

[steady grove]

maam

calm your self

[hoary saddle]

@indigo carbon have you added canny into the AIT workflow yet? getting myself too confused to make it work lol

[vital ermine]

Okay little boy you can have the nickle candy.

[hardy cipher]

oh, I think that might be close to if not exactly the sample workflow, just spread out. I've had that error a couple times myself. it's kind of a beast and seems like it's probably not exactly the workflow itself. could be wrong though

[wet nacelle]

Do you know how to fix it?

[hardy cipher]

hmm, let me look at it. I have it logged somewhere

what's the bottom of it say?

[wet nacelle]

Hang on sorry. I closed it out. I'm waiting for the error again.

[hardy cipher]

if you get it again, maybe paste the last several lines. and the first few lines as well if possible. but usually it's the last few lines that are the real source of the error

[vital ermine]

Thanks for finally adressing the issue.

[visual glade]

if you update that should be fixed

[wet nacelle]

As in the secondary updater?

[visual glade]

update/update_comfyui.bat

[vital ermine]

@visual glade I love iA3 as it does everything I need and fast too, and small, so what is a viable alternative (especially when doing style trainings)?

[hardy cipher]

lol, well there you go

[steady grove]

anything offered by kohya-ss

[vital ermine]

my kohya_ss does iA3

[hardy cipher]

I was thinking something with pytorch maybe, but I'd imagine comfy knows more about his own program

[hardy cipher]

are you using the portable version? do you have the manager?

[steady grove]

anything else i think was the obvious assumption. i thought you liked jumping to conclusions?

[vital ermine]

the worst out of all of them was DyLoRA. Man, that thing was slow

[wet nacelle]

I'm not sure it worked.

[wet nacelle]

yes and yes

[steady grove]

dont use the worst one then

[hardy cipher]

if you have the manager, update from within the manager, or try to, then restart the program and also refresh your browser

if that doesn't work there are update .bat files in the update folder in your install

[visual glade]

windows....

can you just move your input folder temporarily then update then put it back?

[shy kelp]

try this, delete comfyui manager folder in /custom_nodes, and reinstall using cmd

[hardy cipher]

you know, I am not the actual expert on any of this, just giving you advice that's worked for me. so listen to the people that know all the things

and don't want to bombard you with 50 different suggestions

[indigo carbon]

not yet, if it doesn't require to compile new modules- I'll start working on it today

[hoary saddle]

appreciate ya, will keep trying things (maybe i'll get lucky) lol

[safe stump]

Hi i have a question. Can i use stable diffusion model in my app / platform

[hardy cipher]

best thing I did for myself in regards to comfy was a completely fresh install

for whatever reason things really got bogged down in the last one

[indigo carbon]

link to Cnet? I haven't used it yet so I don't have it on my PC

[hoary saddle]

https://huggingface.co/stabilityai/control-lora/tree/main/control-LoRAs-rank256

https://huggingface.co/diffusers/controlnet-canny-sdxl-1.0/tree/main

not really sure exactly which to use, get somewhat similar results using the control loras in the controlnet model loader as i do with the diffuser bin file

[indigo carbon]

I'll test this one

[wet nacelle]

@visual gladeThe python update then base update within the manger seemed to have worked.

[steady grove]

https://huggingface.co/thibaud/controlnet-openpose-sdxl-1.0 too

[steel mist]

This is going to sound like a cop out but you should review the license agreement with a lawyer who’s familiar with your jurisdiction https://github.com/CompVis/stable-diffusion/blob/main/LICENSE

[tepid maple]

or use chatgpt to read for you and tell you what you can or not do

[steady grove]

the short answer is yes, but if this guy is worried he should really do a full review of the permissive licenses

[tepid maple]

🙂

[steel mist]

Also uh how to I collapse that stupid preview

[steady grove]

lil x that shows up top right of embed

[steel mist]

Hrm don’t see it on the iPad app

[steady grove]

dont know bout that one

its a mouse over thing on desk

[steel mist]

Also FWIW I’ve had really mixed results (compared to canny/depth) with that open pose model

[steady grove]

its behaved like a typical open pose model for me. which will be very different from depth models

[steel mist]

I was getting extra limbs and stuff which is exactly what I would expect openpose to not do

Maybe my expectations are too high… never used the 1.5 model

[daring smelt]

maybe don't say things like this and expect people to engage with you in a constructive way

[safe stump]

Thanks mate

[vital ermine]

I asked so many times, or brought it up but never was answered. At least now I have a real answer so all is well.

[wet nacelle]

@hardy cipherIs there a way to denoise the canny image so that the base image is more present? I can only use "strength" and that's not doing what I thought it would.

[hardy cipher]

decrease denoise in the sampler

[wet nacelle]

ah

[daring smelt]

Understand the frustration, but not cool to degrade the very devs you're so desiring an answer from

[soft zealot]

and bear in mind thats only the Licence for the OG SAI Model. Subsequent models derived or fine tuned from the base model may have their own licence restrictions over and above the base restrictions of this licence whch are th eminimum requrirement sfor any derirative.

[void zenith]

yo how to quit watching anime girls

I will read easypeasy methosd

hope it works bro

[steel mist]

Yeah for sure, it’s definitely “talk to a lawyer” territory if you’re looking at making any significant (monetary or time) investment in an app

[void zenith]

ok ty bro

[wet nacelle]

shut up

[soft zealot]

Thats an intersting section of the licence
"To defame, disparage or otherwise harass others;"

Technically that means you can't use the model to produce Memes extrating the urine out of people ;o)

[hardy cipher]

no deepfaking and harrassing others with them

[steady grove]

I tried the most gauntlet test i could think of for the thibaud model. and this is literally just my first one off. Set it up, typed a cheese prompt, fired it off. Why are people hating on this?

[vital ermine]

Where did I degrade a dev? What I was saying is I have asked. I have asked directly, indirectly, to comfy, joe, and I am ignored, or was. Not much else can I say about it so there was my frustration by being ignored YET other people asking other question before, and after, are freely answered. If that doesn't feel a bit discerning to anyone I don't know what else would. As I said I finally received the answer I have been trying to get for a long while now.

[hardy cipher]

people hate on everything

[daring smelt]

Silence is not golden when it is asked and the devs go deaf, dumb, and mute.

[hardy cipher]

the people that hate on it don't limit their hate to one or two things. it becomes habit and they need it to feel complete

[steel mist]

Specifically, I was having trouble with any model that had crossed limbs, even with a single person. Using the open poses model. I don’t want to hate, but I was not getting consistent results like I do with canny/depth

[steady grove]

Last night i shared this and a Voldemort showed up and said it was trained on a cheap gpu with 100 images and sucked. like.. .how are people concluding these things?

[safe stump]

Damn how do I get all of this I have an idea for a platform so I want to use a stable diffusion model for some features

[soft zealot]

if in doubt the answer is 42

[hardy cipher]

even if I do think something does kind of suck, I just remind myself it's probably not for me and let other people enjoy it if it's what they like

there is literally no benefit to behaving like that

for anyone involved

[steady grove]

Just coming in here with made up nonsense like it's a fact. i wonder if there's a campaign to cause disinformation anywhere possible in order to slow the proliferation of open sourced AI. Now THATS a conspiracy theory.

[strong field]

@visual glade , got an opinion on AIT vs TensorRT and which one will be the more practical solution moving forward?

[hardy cipher]

that sort of attitude is indicative of a person that's not happy with their own life

[steady grove]

I think the real problem is the preprocessors we have available in comfyui. they are producing some jank pose estimations for me.

[tepid surge]

yippe meta data stripping begins

example:

[hardy cipher]

well having criticisms is fine

but some people just hate for the sake of it

[steady grove]

different models for different purposes for sure

[steel mist]

FWIW I was using ones downloaded from openposes.com not rolling my own. Through the pre processors have seemed pretty good?

[vital ermine]

I disagree because those are not derogatory terms those are what happened by being ignored. Not everything is a derogatory term. All they had to do, which Comfy did today, was simply answer the question as to why it was not implemented. The last straw was Joe ignoring my DM he asked us to give for suggestions. Not nice being ignored and feeling marginalized.

[peak dove]

ComfyUI

[daring smelt]

dumb is not positive connotation

[steel mist]

Anyways I set it aside because I didn’t want to figure out the root cause when canny/depth were mostly doing what I wanted

[steady grove]

here's the estimation i caught from the group photo. good but not super great

more open pose models will show up in time no doubt. probably even a control lora for one. but i'll use this for now. it works great.

[vital ermine]

Oh, you thought I meant IQ? No no dumb "deaf, dumb, and mute". Not the colloquial term for IQ rather the real definition "temporarily unable or unwilling to speak."

[hardy cipher]

people also forget the outlandishly fast speed everything has progressed

think of text to image ai 1 year ago

what else has progressed this quickly

[steel mist]

I’d love to see some winners if you’re OK sharing the workflow. I dunno why I was getting janky results

[hardy cipher]

in the history of anything? lol

[steady grove]

the medical term, like ritard, is why people started calling each other dumb. Because of that, medicine has moved on from that word as it has a negative connotation. Mute is the new term. It's not "dumb and mute". It's "Mute used to be dumb but that changed because of the aggression behind the old term"

the mod is giving you good advice

[strong field]

its painful watching a train wreck in real time

[steady grove]

actually really love super 8 movie for that

[strong field]

shit, great reference, that was an amazing sequence

[tepid maple]

https://tenor.com/view/chernobyl-not-great-not-terrible-average-gif-14312851

[steady grove]

https://tenor.com/view/nuke-gif-8044239

[steady grove]

my workflows are atotal mess haha. for controlnet part its just the typical hookup. Apply ControlNet (Advanced) and is the keystone i'd say. sits between the clip conditioning and the sampler

[tepid maple]

Ijust want to open sd prompt sharing site

most of the prompt sharing sites feels like soup

no order

[steady grove]

I think that sequence is probably abrams best work, but it's really hard to pick a best. he's a hit an miss kind of director, but when he hits it hits

[tepid maple]

no proper tagging category filtering....

even if there is tagging ux is bad or slow

[tepid surge]

outside of the convo context, mm soup

[indigo carbon]

it does require new modules. currently compiling some for testing.. I'd say about 1 week to be added to the custom node.

[vital wraith]

I find prompting to vary a lot between models of the same SD version…

[tepid maple]

couple mothns ago I even coded couple component for my sites even

it is private library but it is possible to make it public site

wrote a prompt extracting script, cleaning script for auto tagging....

[soft zealot]

all I would ask people like @visual glade etc is dont forget the bread & butter users and take development down a path that may exclude some current users.

[safe stump]

What is the true meaning of open source

[spring fulcrum]

Has anyone wondered why models are never really bigger than 6.5GB? I am curious about that... If anyone knows, let me know please...

[tepid maple]

opensoruce is a method

[steel mist]

It doesn’t have a single legal definition

[tepid maple]

not a religion

[steel mist]

Which is why license agreements exist 🙂

[steady grove]

truth is subjective

[vital wraith]

To crush your enemies, see them driven before you, and to hear the lamentations of their women.

[tepid maple]

but opensource works even in gov...

[strong field]

hehe, thats why im asking, as architectures update, and we move forward at breakneck speeds, where are they going to focus?

[queen adder]

My first 10 epochs or Lora training have been..... interesting 🥴 https://imgur.com/a/KHIsCnw

i thought by providing a dataset with a face from a diversity of angles it would help the model learn, but i think it's just confusing it more.

[indigo carbon]

TensorRT is less efficient and isn't capable of moving from CPU RAM to VRAM and back due to using ONNX. AIT can, plus- easier to compile and slightly faster

[vital wraith]

Is that from over baking?

[strong field]

oof should have just asked you mate

[queen adder]

i'm only using 10 repeats so i don't think it's overtraining, and learning rate is 0.0004

[strong field]

@indigo carbon ,Is TensorRT basically limited by ONNX then? Been trying to read up on it and although i see conversion is done to and from ONNX i wasnt sure if you could forego ONNX to begin with

[safe stump]

I thought because stable diffusion is open source i can use it for develop something using it with all those licenses seems like not possible

[vital wraith]

Hm, I got really good results from a varied data set in 1.5. I’d think it would be the same for SDXL. My LoRa just finished last night, about to find out soon myself. Lowered mine to 5 repeats and 5 epochs.

[safe stump]

Better pay for that evil corporation open ai🤣

[queen adder]

100 images x 10 repeats training against 1000 regularisation images. the reason i picked 10 repeats is i think you want the number to be close to the times it looks at regularisation images, a balance.

[steel mist]

Just because there are a lot of licenses doesn’t mean you’re going to be restricted. Licenses can specifically grant rights in addition to restricting them

[steady grove]

openai actually releases a lot. the clip L model that sdxl and sd15 use is a release from openAI

[indigo carbon]

yes, if it used a better and more friendly language like MSVC (what AIT uses..) it would be capable of moving from CPU RAM to VRAM. in that case; AIT would still be superior due to being flexible enough for no need to specify model's shapes. however: I'm not underestimating NVIDIA.. I'm sure if they would try hard enough they would make something better than AIT

[soft zealot]

I do wonder if it would be helpful is SAI came out with sa statement requesting that all developers ensure that there is a "minimum spec" that they build too so everyone know where they stand.

BYRW I am cogniscent that this is very much chicken & egg abd that YMMV but...................

[vital wraith]

I haven’t properly utilized regularization yet, I always skipped it for 1.5. I don’t know how important it is now, soon to find out when I get behind a computer to test.

[safe stump]

What if I accidentally broke the agreement

[steel mist]

Well you could get sued

[safe stump]

Exactly

[queen adder]

I got another question. say im using training images all of the same person. do i have to caption their name in each image, or is implied when i used the name as the class/instance prompt?

[steel mist]

But, and I cannot stress this enough, it’s pretty hard to accidentally violate a license if you are speaking to a lawyer

Good council will give you cover

[steady grove]

negligence can happen in any endeavor. "I could accidentally break the license" is something that business entrepreneurs have been dealing with for 100 years

[vital wraith]

Depends if their name is already a used caption. Often times if it is, it can either cause your generation to vary, or, even look more accurate.

[steel mist]

Also at least in the US you can structure your company to protect yourself from suits

[indigo carbon]

basically: if NVIDIA wasn't so lazy, they likely would make something better than AIT imo

[steady grove]

yeah. nvidia . lazy...

[safe stump]

How

[vital wraith]

I am making a celebrity Lora and tested to see what SD and SDXL recognized before naming my Lora/caption

[indigo carbon]

not only lazy; also cheap.. they could have given the 4000 series good VRAM, but they didn't.. why? to save a few extra bucks

[safe stump]

I just want to make something with my friends doesn't have a leverage to hire a lawyer and stuff

[steel mist]

I am not a lawyer and this is not legal advice, but:

• Setup a limited liability company/partnership. These are relatively cheap to do and the paperwork can be boilerplate.
• Follow the processes that your jurisdiction manadates for that, eg something like a classified ad in a paper of record
• Open a business banking account
• Ensure you are not commingling personal and business funds as this is the #1 way that your liability shield

[safe stump]

It's easy to use dell e with that money

[steel mist]

DallE also has license agreements, costs money, etc

[vital wraith]

Doesn’t an increase in VRAM impact the bandwidth you can read and write to it? I figured it was a give and take situation.

[safe stump]

Oh i know that thanks anyway

[steel mist]

If you’re just screwing around with friends and will not be selling your software or selling access to it, it’s a hobby not a business and you’re probably* (again, not a lawyer) fine

But if you have dreams of making money with the thing you’re working on, upfront structure is your friend

especially if you’re working with friends

[steel mist]

Anyways I work for a legal tech startup so I am pretty conservative on this stuff

[steady grove]

my 4080 has gddr6x . that's not good?

[vital wraith]

I am just plainly fixated on making AI do what I want. GPT and SD heh.

[tepid surge]

that's cool

[safe stump]

Nope we want to make it to business

[steel mist]

Setup an LLP

It’s a few hundred bucks

[steady grove]

i stopped watching youtube hardware channels year ago. they're all so "if its not the BEST, it's the WORST. Don't buy garbage! Like and subscribe!" /waves long hair

[vital wraith]

I used an LLC when contracting to limit my liability. Unless your doing something truely nefarious, it should protect you from legal destruction.

[indigo carbon]

no, it's fine.. but just saying NVIDIA used memory busses with technology from 2013 on them, when your VRAM gets full: it's SO funny to see the speed go to a little over 0 due to the memory bus being so bad

[indigo carbon]

they didn't do that to the 4090 though

[safe stump]

I was thinking c corp because it's easier to raise funding

[steel mist]

Lol if you’re thinking about raises

[steady grove]

might be a configuration error because when my card uses shared memory it works mostly fine and well over 0

[steel mist]

definitely get a lawyer

[tepid surge]

so floofy

[steel mist]

And a good one because VC will fuck you over hard if you’re not careful

[steady grove]

https://tenor.com/view/waynes-world-not-aware-gif-9701210

[indigo carbon]

then VRAM isn't full.. try loading a 33b LLM on a 4080 and see what happens

[safe stump]

Seems like you have experience in this

[steel mist]

Nah just old and remember reading fast company in the dot com era

[steady grove]

i can see the usage in gpu-z and afterburner. won't argue though. you've told me the card isn't good. i must be getting good usage mistakingly. mb

[vital wraith]

I used to do tech contracting/consulting. LLC is the way to go. You’ll need to draft an agreement between all stake owners though.

[safe stump]

Generative ai art is not the whole platform is just a little cherry on the cake i think go with dell e is the better way

[steady grove]

a 33b loaded onot 16gb is slowed down for much different reasons than a saturated memory bus.

[indigo carbon]

you misunderstood. the VRAM itself is great, but watch what happens when it uses CPU ram.. the memory busses NVIDIA used for most of the 4000 series has capabilities similar to what they used on the cards from 2013-ish

[indigo carbon]

oh 4080 has 16? my bad then

[tepid surge]

"steampunk zombie":

[safe stump]

Oh thanks mate

[indigo carbon]

I was talking about the memory busses they used for the 12GB ones @steady grove

idk about the 16gb ones, haven't heard as much complaints about them

[visual glade]

whichever one works best on windows and lets me swap the weights

[visual glade]

right now AIT seems like the best but it would be easy to switch to TensorRT

[indigo carbon]

TensorRT doesn't let you move the weights, due to ONNX. AIT does

[visual glade]

I know but apparently there is a way but I have not tried it yet, will probably try it very soon

[royal monolith]

Hello all 🙂 Do you know if a controlnet inpaint is available? (i.e: we upload a picture and a mask and the controlnet is applied only in the masked area)

[indigo carbon]

by the way, also speed increase from TRT is lower than AIT. I'm sure NVIDIA can make TRT better than AIT, but they don't seem to care enough.. also harder to ship precompiled TRT modules

also the custom node already works on SDXL+refiner, so I'm not sure if it's needed that @visual glade would add it the main branch

[steady grove]

not for sdxl yet. tooling is still getting rebuilt for the new generation

[visual glade]

yeah the custom node works fine which is why I haven't released my own implementation

[steady grove]

has it stopped breaking controlnet flows? i had that issue. the custom node was installed and i couldn't use the apply controlnet node

[indigo carbon]

only issue with it is batch size, but yeah. the precompiled modules are good enough, I compiled some modules myself and speed is identical to precompiled ones.. the only problem is it gives errors when compiling for batch_size>1

@visual glade where you able to compile for batch sizes higher than 1?

[royal monolith]

ho ok I need to be patient then 🙂

[vital ermine]

Can AIT or TensorRT be used for training?

One would think, or hope, but you never know.

[tepid surge]

no idea i just bake beans

[vital ermine]

baked beans are yummy

[tepid surge]

cx

[indigo carbon]

theoretically, yes

[vital ermine]

I read tensor could but no idea about AIT

[steady grove]

holy shit now i want baked beans fo rlunch

[tepid surge]

lunch beans cx

[indigo carbon]

tensorRT would be harder to infer for training. AIT is much easier to infer

idk about speeds though for training

[strong field]

been working on 30B with 3090, painful

[vale eagle]

After actually tried revision, I must say it is unlimited idea for generation.

[tepid surge]

i'm thinking of coleslaw too 🤤

[vital ermine]

? As I said I know zip about ait and not used it but it looks like you bake the model and for training we aren't changing models so should be perfect for it.

[tepid surge]

No one:
Ai:

[indigo carbon]

still, AIT can be inferred by implementing a few lines of code, idk how TRT is inferred for training though

[tepid surge]

love that top right one

[vital ermine]

Well, I would love to see either in a trainer as that would rock

[indigo carbon]

one thing I can say for sure, you can't train SDXL on TRT due to it not being able to use dynamic shapes

[strong field]

They say it can be inferred after converting the weights to shared object, stil a pita for every implementation

[vital ermine]

Well, ugh

[strong field]

great point

[indigo carbon]

AIT however..

[vital ermine]

I wonder if we will see AIT in a trainer?

kohya ait edition, heh

[strong field]

Its already done with AIT, lol

[indigo carbon]

well, if there are some sort of training workflows on ComfyUI, it's possible as of now.

[vital ermine]

comfyui can train?

[indigo carbon]

that's what I'm asking..

[vital ermine]

Ahhh

I was like yeah, baby

[tepid surge]

cx

[indigo carbon]

huh, there is already an implementation in kohya_ss for something called deepspeed, it's very similar to AIT- both compile MSVC optimized modules. except DeepSpeed is meant for training

[vital ermine]

deepspeed I shy away from as it was hell on windows to install then I found out it isn't for windows and I dropped it. Something I had was complaining about it is the only reason I knew about it.

[tepid surge]

What was the prompt for this one cx

[indigo carbon]

theoretically, it should work fine on windows and give the same boosts that AIT does

[indigo carbon]

might be an issue on kohya_ss's implementation

[vital ermine]

Could be

[stone fossil]

Prompt by Mass. 🙂

[indigo carbon]

I mean, it IS made by Microsoft..

[vital ermine]

Yep

[wet nacelle]

@hardy cipher Hey do you think you can find this real quick?

[hardy cipher]

it's one I made https://github.com/picturesonpictures/comfy_PoP

it's not really required, just makes things more convenient for me

[wet nacelle]

ah okay thanks

[hardy cipher]

might try to put it in the manager soon, but just haven't really got around to it

[steady grove]

canonically this kryptonian would be very powerful since he'd absorb more yellow light

[glad grove]

cant wait for the disney's superman

[tepid surge]

time to go get tea brrb

[tepid surge]

yooo sick aff

[stone fossil]

At your service. 😉

[tepid surge]

intricate fantastical creatures; incomprehensibly beautiful; tangible, structured, futuristic entitites; dripping in 600 ghz ferrofluid; 8k resolution masterpiece; award winning; hyper realistic, ultra-detailed

dank uuu

[glad grove]

the saviour

[tepid surge]

Could super man be strapped to just a normal car and then fly with it like it's a flying car cx

[tepid surge]

Anyone have artists they like use in prompts? i usually just throw "painting" or "Hyperrealism" but am looking of more style x.x

[hardy cipher]

@tepid surge Jackson Pollock and Hans Memling, preferably together

[strong field]

https://supagruen.github.io/StableDiffusion-CheatSheet/

[indigo carbon]

which one's better? I'm testing new upscalers

[glad grove]

first one

[tepid surge]

i cant really tell but i did notice the second one's background is darker

[strong field]

1. has a better composition and coherence
2. is more detailed, defined, and crisp, but as a result shows some artifacting

[indigo carbon]

so 1?

[strong field]

lol, yeah 1

[indigo carbon]

I'm looking for an upscaler that would be best in most scenarios.. idk

[stone fossil]

Mod chip.

[steady grove]

sometimes they do be like that

[tepid surge]

Short black hair in front of colorful maze wall

bot left and right look like real artists posing in front of their art cx

[glad grove]

speaking of real art

[stone fossil]

xD

[indigo carbon]

https://github.com/Phhofm/models/raw/main/4xLSDIRplus/4xLSDIRplus.pth so far, this one is the best I found

[tepid surge]

what a gpu die looks like with the cap off cx

[stone fossil]

Oh yes.

[tepid surge]

monkee

[steady grove]

https://youtu.be/-P28LKWTzrI?t=71 exactly how mythbusters explained it

[strong field]

ill check it out

[strong field]

most luxurious

[stone fossil]

Just put thermal paste.

[glad grove]

real

[glad grove]

sensei 🙇‍♂️

[glad grove]

looks like history channel new show about car collectors

[tepid surge]

the clothing detail and flames so perfect cx

[steady grove]

fingers not so much

[tepid surge]

i didnt even notice 😭

[cosmic compass]

whats the best upscaler for watercolor/art , ersgan 4x ?

[cosmic compass]

is that sdxl base model if so whats the prompt for that style ? looks nice

[cosmic compass]

that looks epic whats the prompt model for that ?

[tepid surge]

For killing.

i didnt have libreOffice on my pc for some reason like what

[vital ermine]

I just installed deepspeed in Linux, set accelerate up with no real documentation, and was firmly met with a ton of errors. Probably the accelerate config asking so many questions for it with absolutely nothing I could find for training SD.

I guess saying to offload to cpu was a bad thing "[WARNING] cpu_adam cuda is missing or is incompatible with installed torch, only cpu ops can be compiled!"

several pages of info shot out at me, lol

[tepid surge]

i'd start with cuda from nvidia

[steady grove]

i mean, the docs are here. not using them doesn't mean they're not there. https://huggingface.co/docs/accelerate/index

going to the project homepage helps

[hoary saddle]

managed to get lora and i2i setup, cnet just gets my brain all messed up lol

[vital ermine]

NotImplementedError: Module [LoRANetwork] is missing the required "forward" function
steps: 0%| | 0/155 [00:00<?, ?it/s]
[16:27:21] ERROR failed (exitcode: 1) local_rank: 0 (pid: 15437) of binary: /mnt/WIN_F/kohya_ss2/venv/bin/python

I guess kohya hasn't implemented what it needs

Yep, it will take Kohya to implement this or just don't use deepspeed.

[tepid surge]

As much as i love linux i'm just not technically savy enough for it cx

[vital ermine]

I disabled it with the accelerate config command

[indigo carbon]

Cnet doesn't work with AIT. but good job, I don't see why you modified how I organized the workflow, but yeah that's the right way to use LoRA and I2I with my AIT workflow.

[hoary saddle]

sorry, i had to move them to see where all the lines went and then didn't remember where you had everything lol

[hoary saddle]

ah so cnet doesnt work with AIT, probably why i'm having so much trouble, good to know

[indigo carbon]

yeah, new modules should be compiled. next step is to do batch sizes, then we can start to focus on supporting Cnet

[upbeat summit]

nice - 26gb images in my ComfyUI\temp directory. hah - guess it wasn't cleared for a while. what did I break?!

[strong field]

Lol

[hardy cipher]

That's from dreamstudio. Mess with it on my phone sometimes

[hoary saddle]

maybe it's just me but it's not that AIT is faster but the generations seem to be better

[upbeat summit]

that looks pretty good!

[wet nacelle]

I've been hearing about this "AIT" thing. Can you link me a place where I can read about and maybe get it running myself?

[hoary saddle]

i2i, cnet would make these so much better for sure

[strong field]

Unified inference is the name of the game

[hoary saddle]

i installed it from some links @indigo carbon sent and from the comfy manager

[strong field]

Optimized for architecture

[hoary saddle]

than just dragged one of his images into it

[wet nacelle]

Kool. I don't really have the resources though. You really just gave me more info without giving me more info.

[wet nacelle]

I'm just confused

[indigo carbon]

if your GPU isn't at least 3000 series, unlikely to expect the massive boost other users report

[strong field]

https://github.com/FizzleDorf/AIT

[wet nacelle]

Kool. I have a 3070Ti.

Am I just going to be told what I need without being told where I can find it?

[wet nacelle]

Thank you.

[indigo carbon]

if you have ComfyUI manager you can just drop on of my gens and click "install missing nodes"

[wet nacelle]

Ah okay. Thank you.

[steady grove]

tried your workflow on a favorite image of mine and it was so nice. good params thanks for sharing

[tepid surge]

this underwater stuff so cool

[tepid surge]

was harder to get hispanic, pulled up more line art and black and white then previous xnx

[cosmic stratus]

Hi, trained a SDXL Lora using https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/README_sdxl.md it worked however the resulting safetensor is not compatible in A1111, anyone knows why that might be and how to convert it?

how is everyone training SDXL Loras these days?

[vital ermine]

@indigo carbon Apparently deepspeed can't work with loras? https://github.com/kohya-ss/sd-scripts/issues/711

[indigo carbon]

@visual glade did you run into any issues when compiling AIT modules for batch sizes? me and a few others tried it today and it doesn't compile for batch sizes overt 1

[visual glade]

yes

[vital wraith]

Was Nvidia high when they priced the 4000 series? Why would I pay this much if I already have a 3000. O.o

[indigo carbon]

can you share the code you used to compile for batch sizes? this is where we are stuck with the node

[cosmic stratus]

@vital ermine what are ppl using to train Loras for SDXL?

[vital ermine]

Kohya

[cosmic stratus]

thank you, I'll try that

[vital ermine]

Be prepared it requires hefty gpus

[visual glade]

my code isn't compatible with the node

[cosmic stratus]

3090 wouldnt' do?

[vital ermine]

perfect

[visual glade]

it uses the same key names as the current model

the node uses diffusers key names

[indigo carbon]

when compiling with scripts/compile_sdxl.py with *--batch-size* flag it doesn't compile. did you make your own compilation scripts for this?

[indigo carbon]

why not, doesn't it output in .dll or .SO format?

[glad fulcrum]

Loras work on auto1111 as far as I know.

[strong field]

Are the key names easy to convert to current model? Assuming you rewrote compile scripts

But I didn’t realize this may break node functionality

[indigo carbon]

if the code he made is really capable or compiling with --batch-size he must have changed something... however, the modules are the same format. if @visual glade would ship AIT nodes there is no reason they shouldn't work with the node unless comfy changed something in the profiling

[strong field]

@indigo carbon this might mean we are limited by hlky code if we want to expand the module libraries

[cosmic stratus]

it's some issue with loras trained using diffusers not working with a1111

it's exactly this issue
https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/12448

it's closed but the solution doens't actually work for sdxl loras

[strong field]

You or comfy may need to correct me if I say something stupid but it doesn’t sound like comfy used the hlky code to compile, in which case his AIT implementation wouldn’t be compatible. We would need to rewrite hlky code if comfy passes the modules

[tepid surge]

niiiice!

[strong field]

My question is, is it as simple as rewriting the module json, or if it requires rewriting AIT scripts

[indigo carbon]

that isn't likely at all. if the AIT module is in a .dll and/or .SO there is no reason for it not to be compatible

[indigo carbon]

@strong field let's go to DM, looks like there is work that need to be figured out

[vital ermine]

lol

[icy brook]

Thanks!

[vital ermine]

even better on the new model

I can't wait for more models to arrive

Interesting merge of two loras and the new model

[hoary saddle]

was all @indigo carbon, i'm just bumbling around making a mess of it 🙂

[tepid surge]

is that related to the AI template thing that was talked about earlier, I only half payed attetion then x.x

[tepid surge]

Dem clouds

I got a can of canned fruit the other day, and it was the right purchasing decision om nom

[vital ermine]

No, this was some new model released on civit yesterday.

[tepid surge]

ah, 💯

[glad grove]

is it uncensored?

[vital ermine]

I do not know

mostly women in it but above and some men I did get too

[soft zealot]

@indigo carbon ok soI think I've compiled the basic AIT stuff for use with my 1080ri (you never know until you try lol)

Doyou have a basixc workslow with AIT enabled bits in I could use as an initial tester please (or have i misunderstood somehwere [entirely possible]?)

[tepid surge]

i added angry instead of smiling 😭

[indigo carbon]

you misunderstood. when leaving an AIT module enabled, it will try to apply the module on the model it's using, should be enabled only on workflows without refiner... load one of my recent gens, they have the workflow in metadata.

[soft zealot]

ah Ok Im a great big fat donkey then

I went ahead and did some guuff from here lol
https://github.com/facebookincubator/AITemplate

[tepid surge]

aaaaaaaaaaaaaaaaaaaaaaaaaa

[indigo carbon]

you compiled SDXL with that? the code in that repo isn't capable of compiling SDXL as far as I know. you need Hlky's repo for SDXL

[indigo carbon]

OH

I didn't see that

yeah you should still use Hlky's repo

[soft zealot]

and then grab one your workflows?

KK

Lets go find that other repo

I';ll just leave this one to dfester lol

[indigo carbon]

still no. you need to compile the AIT module, not the build module. DM me if you get stuck on something

[visual glade]

the problem with the current AIT nodes is that it needs to convert a bunch of keys because the modules use diffusers key names

in my modules I made it use ldm keynames so there's no need for conversion

[tepid surge]

Time for animation stuff

[wet nacelle]

@indigo carbonYour workflow makes my PSU sounds like a rather quite fax machine.

[tepid surge]

fucking mtb, changes a property to an input node but doesnt add a node that's accepted by it, please shoot me

[glad grove]

least edgy R6 player

[tepid surge]

ooh my goodness. i got it work. I slapped a wildcard node on it

[glad grove]

sensei

[wet nacelle]

ahoot

shoot

[shy kelp]

how do you made this? so amazing

[hardy cipher]

Pretty boring prompt really "Steven seagal is a futuristic tangerine dragon dripping in 600 ghz ferrofluid; 8k resolution masterpiece; award winning; hyper realistic, ultra-detailed; vivid, epic, exhaustively detailed and realistic, ultra high resolution masterpiece; Steven Seagal face"

[glad grove]

https://tenor.com/view/dono-wall-talking-wall-bricks-gif-17741481

[wet nacelle]

yes

[shy kelp]

a1111?

[hardy cipher]

Nah, did that one on dreamstudio. It cost me almost 2 cents to make

[hardy cipher]

VOLDE ADY DUST

[tepid surge]

holy guac please let me in on your positive prompt

[wet nacelle]

We trust

[glad grove]

whose dad is this

[hardy cipher]

got a little something going on in his bathing suit area

[shy kelp]

portrait of the most beautiful form of chaos, elegant, a brutalist designed, vivid colours, romanticism, by james jean, roby dwi antono, ross tran, francis bacon, michal mraz, adrian ghenie, petra cortright, gerhard richter, takato yamamoto, ashley wood, an evil girl wearing intricate dead details flowers, intricate flowers as dress, ((decay and spooky)), plants and branches, atmospheric, trending on artstation. 8 k masterpiece

quite long prompt

[tepid surge]

ooo beautiful, all those artist too xox thank you, will investigate

i'm in a graph training montage rn cx
https://youtu.be/maRVwVpSCQQ

https://tenor.com/view/weight-lifting-workout-hot-girl-cleavage-abs-gif-18591080

[hardy cipher]

decent form

[glad grove]

feelin kinda hot rn ngl

[upbeat summit]

intense

[hardy cipher]

tried to capture charles bronson wrestling a monkey in an analog black and white image 🫤

[tepid surge]

Me after cramming 40 min of random ML info into my brain:

[queen adder]

will instance prompt recognize two words or do you have to use one?

[hardy cipher]

I guess you can try and find out

[glad fulcrum]

It seems that yes, because of this:
https://www.youtube.com/watch?v=N_zhQSx2Q3c

[hardy cipher]

I wonder if the person that came up with the refiner model idea is sad people don't like it

[queen adder]

thanks!

[steel mist]

SAI seemed really excited about it, and the paper talked it up a lot. It's surprising to see it sort of cast aside

I wonder if it's because SAI has to use a more limited training set than we do for finetines

[hardy cipher]

yeah, just thought it was weird how it kind of went from "woah! check this out!" to being put on the back burner

I guess it's not going away though

it does add complexity though, and I don't know if anyone really know hows to use it right

[azure oxide]

i saw a few devs mention on release date that they dont expect it to be necessary going forward

[steel mist]

My guess is there's just a lot of low hanging fruit with base

[hardy cipher]

I don't hate it, but I cannot get in the groove with it. I always question what I'm doing with it

[tepid surge]

idk what happened:

[steel mist]

and once that get exhausted people will start doing refiner fine tunes as well

That and better samplers in comfy, etc

Like whatever foooooOoOOOcus is doing with the "momentum" sampler

[hardy cipher]

it might end up getting used more with 1.5 models honestly. seems like it does more for them

[tepid surge]

I'm testing CuteLora on CivAi

[hardy cipher]

those are pretty cool actually. not sure what the hell is going on, lol

her eye looks painful

[tepid surge]

yeee

Without lora Left, with right: (same prompt, seed, sampler, etc)

I literlly couldnt tell the diff other then the lips

[tepid surge]

Oil painting lora:

[queen adder]

i've tried training a model on UFC fighters before but i could never get the model to differentiate fighter A from fighter B and the limbs would be all messed up

[hardy cipher]

it thinks they're one big mass of muscles and limbs

[hardy cipher]

whoever said that they had to push clip vision strength up to 2 or 3 for it to work is a madman

it makes garbage for me at 1.7. not sure how 2 or 3 could even work

[tepid surge]

for LoRAs?

[hardy cipher]

no, unclip conditioning

with clip vision

[tepid surge]

wooosh that just went over my head cx

You're under arrest

[hardy cipher]

in comfy. it's a way of essentially using images as part of your prompt, turns them into conditioning data

it can make very cool, and very off the wall things

[tepid surge]

oooh yeah i think i remeber seeing that in Mack's graph

[hardy cipher]

also, it can go very wrong if not balanced correctly

I'm working on a workflow right now that I haven't quit got the tail end of

[hardy cipher]

basically 2 separate prompts and samplers, then send the resulting images through clip vision to make a combo image in the end

[tepid surge]

That's cool

[hardy cipher]

but need to switch something up. it's not cooperating very well at the moment

firs two images come out fine, but then it all falls apart at clip vision, lol

but I've been making some pretty neat things with it

[tepid surge]

i'm just throwing LoRA's at the wall to see what they do till i'm ready to figure out AITemplate, casue apperntly it makes generations go zooom

[hardy cipher]

you gotta use my 10 lora loader

[tepid surge]

idek how to use one yet

[hardy cipher]

hmm, well what's your issue?

[tepid surge]

They're fine >u< i'm just finding they're more specifc, bbl! x.x

[craggy ibex]

blue hair, white shirt, blue pant, red cap with blue brim, male, 1man

[warm beacon]

hello

[hardy cipher]

yes, hello

[hardy cipher]

sometimes when gpt-4 refuses to give me the information I want I call it names like "dingus" and tell it that it's gate keeping, and then it gives me what I want.

fyi, it's a solid approach

[tepid surge]

https://youtu.be/JOCW54f0BHc

[vital wraith]

Do you find you need a higher CFG with XL than 1.5?

[hardy cipher]

for me it's the opposite

but maybe because I'm using different tools

[tepid surge]

I forgot about CFG I usually have mine at 8

[tepid surge]

so fkn over the top cx

[tepid surge]

I knew snoop dogg was a mason

[vital ermine]

Hehehe

[tepid surge]

haya~!

[tepid surge]

oh shit. cx

[mild path]

I knew he was alive

[steel mist]

this is probably the dumbest/messiest way to do this, but figured I'd actually do some testing before I continue to dump on the open pose model

What's the most "fair" way to test these? Right now I have strength at .8 and end % .8

should I just go full bore 100% across the board?

Only so many variables I can mess with and have 3 sets of prompts ⚰️

[steel mist]

and a dozen images

[tepid surge]

cx

[upbeat summit]

I guess you pushed the style by a lot hehe

[tepid surge]

the prompt:
American, wall as background, sunlight, eye shadow, highly detailed, dynamic lighting, vibrant, happy, (hyper realism), neon lights, smiling, painting

cx

i was literlly like "i wonder what'll happen if i just put 'american'"

[vital ermine]

what I got with that prompt and my lora

No, none of that in that pic did I train on, lol

[tepid surge]

cx only darkness and glowing windows

[vital ermine]

I changed the seed

[tepid surge]

dune nose tuuube xwx

i need to learn dark prompts

[vital ermine]

Interesting but I went back to the other seed and removed painting (trained on real world stuff)

dune nose tuuube xwx

Some odd reason the nose tube is mia

[steel mist]

me two weeks ago: "who are these people who put 'watermark' in their negative prompt seriously"

me today: ugh I forgot to add good negatives

[tepid surge]

My age of darkness begins

[half ivy]

[tepid surge]

(right? life turned up to max cx)

[tepid surge]

cw: blood

[glad grove]

wdf goin on in atlanta

[vital ermine]

Just an average day.

[tepid surge]

god damn it!

cx

you beat me to it

[tepid surge]

was gonna send this:
just another day
cw: blood

[vital ermine]

bball, the darkside.

[tepid surge]

i cant post a lot of these cx

[vital ermine]

Dang, dig this

That aint human

[tepid surge]

ooo blade runner vibes

[tepid surge]

xD the glass

[vital ermine]

🙂

[steel mist]

So I'm not sure this teaches me anything except I found a seed that the depth processor does not like but

1. Canny is amazing
2. midas is solid but falls apart situationally
3. open pose generates nightmare fuel with complex poses

[tepid surge]

cw: gore like

[steel mist]

and also trashes the prompt

I am genuinely impressed with how well OP recreates the general pose/attitude based off stick figures

with the two dudes

[tepid surge]

x.x

[tepid surge]

roswell in the forst? x.x

~

[tepid surge]

Okay i think I might be done image gening for a while, binge has been great though xwx, sd xl is 👍
(commemoration of joy & friendship:)
https://youtu.be/G9UA479grns

[peak dove]

MJ has InPainting?

[hardy cipher]

I don't know

wtf is pooled output?

it's really doing me dirty

[dense chasm]

https://www.youtube.com/watch?v=I14b-C67EXY this guy is long live than paul allen and steve jobs, insane and be passionate is not the only answer

1980 he promote mircrosoft software since

[hardy cipher]

lol. he's a goof. but can't really be mad at him for that

[dense chasm]

yup, so the crown should be longlived than batman?

maybe that's the reason why so many ppl in favor of the joker

[hardy cipher]

when I was in high school we took a tour of the microsoft campus for honor society. the guy that gave us the tour said he started working with microsoft when they were still working out of a garage. didn't think much of it at the time, but that had to be one of the first 10 employees or something

I've always wondered if it was actually paul allen. he was pretty goofy

[dense chasm]

yup,so Google , elon musk come afterwards

[hardy cipher]

elon musk lost his mind

[dense chasm]

vice captain is a good position, the caribbean pirates caption will return change again loll,ballmer is clever

never mind,this vice president will be in control,ridiculous

riddleman is no sense,joker has weakness all the time,batman and catwoman in love,the left is in charge during the choas Imao

[shy kelp]

does comfy ui have acces --highvram?

[dense chasm]

Elon is ambitious and worldwide,why he divorced so many times,he is a revolutionized guy,be respect,the point is the veteran internet of PC,like Bill Gates,he is an awesome,long live than steve jobs,during their period,gates is clever dude

[hardy cipher]

yeah, definitely not dumb. elon musk is obviously smart too. but I think he has some issues. literally has to have people looking at him all the time. bought twitter to burn it down. good stuff

[shy kelp]

canny

[dense chasm]

maybe Elon not only buy twitter,maybe some prority in his purchase list,cuz behind twitter,there are a bunch of unvisible boss,when elon choose manufacturing,it's a different supply of chians, so mutual benefits

[hardy cipher]

I don't know. seems like he got rid of a lot of their top tier employees

figured out the node, lol

it wasn't showing up in the errors at all. no reference to the node pack. but then it dawned on me

lol, such beauty

[uncut fiber]

yes he is well made in SD 🙂

[soft zealot]

everyone loves a picture of a cute fluffy bunny wabbit don't they?

and everyone loves a girl in a swimsuit. But what about a rabbit in a swimsuit?

[hardy cipher]

a year ago this would

[rugged crow]

is this a thing?

[vale eagle]

I am trying to use your workflow from image But it is missing Seed With Text and Int to Text. Where could I get that?

[uncut fiber]

Not sure it is so much. Loading model is probably biggest issue in those times.

this recall me to test coloring things

[hallow summit]

when you measure the time from launching the UI to getting the image, or change only part of generation parameters (like refiner seed) - it might be something like that. Comfy starts pretty quickly, and re-uses results of computations that were already done once for given input (so if it doesn't re-generate output from base model, if input for base didn't change, for example)

[shy kelp]

yep

[soft zealot]

theres a link to it in Credits & Notes

[craggy ibex]

What prompt allow checkpoint to generate only middle shot?... hmm

Half body view?

[vale eagle]

Do you mean civitai? I just downloaded your image and drop to comfyui

[soft zealot]

look in my workflow in the Credits & notes

[west breach]

cowboy shot?

[uncut steeple]

[vale eagle]

Found it. Thanks.

[craggy ibex]

Bad Eye

Myeh, I can fix that

[indigo carbon]

Is that a good thing or a bad thing? people that have coil whine report the coil whine to be higher pitched when using AIT

[craggy ibex]

Which one do you like

1 or 2

[soft zealot]

Depends
@vale eagle
You can jumper or unjumper the enablers to suit , Revision can be turned on/off by setting to an approriate value, same goes for controlnet

Wabbits Assemble!!

[icy brook]

Is there a free inpainter online that works well on 1024px images?

[craggy ibex]

why 1024x1024 not work for me?

[uncut fiber]

they almost wear proper costumes 🙂

@craggy ibex what model are you using? Are you sure it is SDXL?

[craggy ibex]

]

[craggy ibex]

Sorry I don't use SDXL

[uncut steeple]

This is the SDXL chat

[craggy ibex]

There are no any SDXL that can generate anime style

[uncut fiber]

then 1024x1024 will not work for you

[craggy ibex]

Oh ok

[craggy ibex]

Where can I discuss about the og stable diffusion then?

I thought SDXL and SD are the same thing

[uncut steeple]

#1072238304042438758

[craggy ibex]

ty

[dusk mica]

Hwo do you prompt sdxl, full sentences or several words sperarated by comma?

[uncut steeple]

Also there are, check civitai

[noble shoal]

@visual glade is it possible to implement, that a muted reroute node breaks the processing chain? That would help, instead of hand wiring switches.

[uncut steeple]

A mix of both works well for me

[craggy ibex]

sadly in #1072238304042438758

We can't send images

[uncut steeple]

#🏞｜general-with-images for now then, ill bring that up

[craggy ibex]

Oki