✨｜sdxl

[soft zealot]

feel free to grab any of my images however Im not expecting anyone to suggest improvements ta

[indigo carbon]

I don't know what "diown" means, but I'm assuming you meant down. and well, that's just how the model was built. it isn't intended to get empty latent images, the base can.

[soft zealot]

it works for me

[uncut steeple]

Guys relax, there are different approaches

[soft zealot]

99% of the time I am not using SDXL1 base & SDXL1 refiner

I am using Model1 ^& Model2 where that could be f say ProtoVision as Model1 & DynaVision as Model2

Hence why I think og f this as a multisampler approach rather then a rigidly defined base & refiner (even though its based on that flow originally)

[uncut steeple]

Same for me right now

[soft zealot]

and actually I think SDXL0.9 was actually superior to what was released as SDXL1 but thats just a personal onion

[uncut steeple]

Im using dreamshaperXL for my "base"

[indigo carbon]

personally; I don't have a personal onion, stuff's making me cry

[hardy cipher]

automatic image changer upper

[soft zealot]

Enchanted Dark satanic Ritual holding a Mobius strip, Auburn Fishtail braid hairstyle, Clear skies, Sketched, Pixiecore, volumetric lighting, (art by Charles Maurice Detmold:0.7)

Image 1 SDXL1 Base & Refiner
Image 2 SDXL0.9 Base & Refiner

Ok bit abstract and maybe not the best example but.............................

[royal python]

It's so weird how I can't get the likeness of someone right in SDXL, while with SD 1.5 it was so easy

[hardy cipher]

what's so weird about that?

[royal python]

That I can't get the likeness right

[hardy cipher]

of who?

[royal python]

of anyone

[hardy cipher]

I could type me in there. "get my likeness" it wouldn't do a good job

[royal python]

with 1.5 I could start to see the likeness just at 500 steps, at least signs of it

I'm talking about training

[hardy cipher]

oh, training?

[royal python]

yeah

I've tried like every setting someone has posted here lol

maybe something is wrong with my PC

[hardy cipher]

yeah, what method are you using? I don't have much experience with training tbh. not with sdxl. probably did 100+ models and loras with 1.5

doubt it's your pc unless you're seeing memory issues or something

errors

[royal python]

yeah SDXL loras basically, oh well I'll keep trying different settings and captions etc lol

[hardy cipher]

all I can say is keep trying. a lot of what people have to say about things like that aren't opinions masquerading as facts. some people know a lot more than others, but I wouldn't suggest taking any of it as the gospel truth. most of this stuff isn't an exact science anyway

doppelganger machine

should definitely turn denoising up

90+

[royal python]

yeah we are all still in experimental mode, no 100% way really

[uneven gazelle]

Hey guys! I'm trying for the first time Deforum extension in Stable Diffusion following a youtube tutorial. When i press the generate button, i got this error message :
"Error: 'Invalid control character at: line 2 column 632 (char 633)'. Please, check your prompts with a JSON validator. Full error message is in your terminal/ cli.
Time taken: 0.0 sec."

Anyone knows how to solve that by chance ? 😅

[hardy cipher]

that's what I like about it. a lot of people want polished and refined, but that just means less to explore

[hardy cipher]

well I'd suggest looking at the full error message in your terminal

says right there to do that

[uneven gazelle]

Yeah right, it tell me to check my prompt with a Json Validator, because i have a "json.decoder.JSONDecodeError: Invalid control character at: line 2 column 632 (char 633)"
But i don't even know how to access my file to see where's the line 2 column 632 and i got no experience in coding in general.. Thanks for the help tho!

[hardy cipher]

well, that has to do with how it's interpreting data. hard to say exactly. it's not so much knowing how to code as it is tracing back where it's coming from

and not sure what you're using for deforum. a1111?

[uneven gazelle]

Ok i think i need to adapt my prompt. Maybe an issue somewhere

I used the same prompt as the youtube tuto and it seems to work! I just need to adapt it with what i want to generate and slowly try to get it!

Thanks for helping tho!

[hardy cipher]

hmm, okay

[uneven gazelle]

Yup i'm using deforum a1111 i guess

[hardy cipher]

well somewhere there's a .json file with some bad data on line 2 in column 632

[uneven gazelle]

Yep that's pretty weird

I think i didnt respected some rules with my prompt there

And so when it converted my prompts as words in JSON format, it could have made some errors or something like that right ?

[lament cave]

hi, is it possible to train SDXL loras locally with 2080ti 11GB?
guess 2x2080ti with NVLink doesn't help?

[hardy cipher]

yeah, things get weird with video. did you check the validator?

[icy brook]

anyone who has solved the disgusting freckles issue that is still here with sdxl? 🙂

[wet nacelle]

sweet

[hardy cipher]

put "disgusting freckles" in the negative prompt

[icy brook]

I meant, when prompting for freckles

I might have to try and finetune for it myself

but rn it's still so extremely bad that I don't know how to tackle it in prompt

[hardy cipher]

put "good freckles" in the positive prompt

[hardy cipher]

auto doppelganger setup

gets a quick danbooru description of the image I can add to, then runs it over the original image with controlnet

could be finetuned a lot better as far as loras and what not

[icy brook]

on Aether Real I sort of can get it ok.

but it's always a lot of it

[wet nacelle]

Looks great

[hardy cipher]

thanks. still need to add my nodes to the manager. I think a couple of them might actually be useful to a couple people

[wet nacelle]

@indigo carbonYou may be the wrong person for this context.

I have some bad news. It's getting better!

[wet nacelle]

true

[indigo carbon]

That's Midjorney?

[wet nacelle]

yeah

[indigo carbon]

Tf?

They made a new version? That ain't 5.2

[wet nacelle]

Did you expect it to not improve?

I'll look at the update channel to see if there is any info on it.

[indigo carbon]

Is this their showcase or is it actual generated with the bots?

[wet nacelle]

Just users showing what they got

more so my friends showing what they got

[indigo carbon]

Damn. I guess we did lose, huh?

If it makes that consistently, it's actually better than SDXL

[wet nacelle]

Not really. It seems that they were just messing around and found a silent update.

Stable also will always be better due to how much we can just flat out do.

[wet nacelle]

the above applies

[indigo carbon]

Idk, it will go silent for a while now due to SDXL be waited on for a while

[wet nacelle]

What does that mean?

"it" as in?

[indigo carbon]

But it is odd that right after SAI made a model that can do text while maintaining quality does make me raise my eyebrows

Same happened right after 1.5 released.

Is MJ finetuning SD?

[wet nacelle]

It's not unlikely. Would they not be allowed to?

In terms of keeping the image of MJ is always private and makes and trains models without outside resources. It is likely that David doesn't even work with the team that much anymore.

[indigo carbon]

Idk, the fact that about a month every time there is a new SD model there is also a new MJ model is very suspect of MJ finetuning SD..

[wet nacelle]

I just think they are.

[indigo carbon]

I honestly don't care anymore, I'm about to take a break anyway, so I don't care if MJ wins for now..

[wet nacelle]

ah okay

[wet nacelle]

https://tenor.com/view/oh-my-god-so-good-bite-hand-gif-14530505

[indigo carbon]

anyways, consider this as goodbye for now. I don't care that MJ is better now; just an excuse for me to go silent for a while..

[wet nacelle]

I feel ya

[uncut steeple]

Poor michael jackson

[crisp owl]

See ya around ✌️
Breaks are good every so often

[wet nacelle]

https://tenor.com/view/mikey-yakson-michael-jackson-yeehee-moonwalk-dance-gif-21514412

[indigo carbon]

nah, it's just futile. every time SD gets a lead on MJ they just release a new version that beats it. I'll probably return next time there is a new model.

[wet nacelle]

dude had his full email

@past oracle

You should change your name.

[crisp owl]

I personally couldn't give two shits about what MJ does.
I can't use it for free on my own pc. I can't tweak every single setting i want.

MJ is like the iPhone compared to Android of AI images to me.

[indigo carbon]

well, consider this as goodbye for now. see ya around guys

[noble shoal]

Close up photo of a corgi holding a sign that says "Bye Tdg8uU" on it. (I am at work and can't generate atm.)

[indigo carbon]

get MJ to do it, they probably can..

[noble shoal]

Hell no.

[wet nacelle]

The MJ community is the only place where I've been able to help others understand how SD is better. I can actually talk to people instead of children. @crisp owl

I don't talk to children often. I talk to people that have the mental age of a child too often.

[crisp owl]

🤣 "dating profile photo of an orc"

[uncut steeple]

[uncut fiber]

This W. T. Pooh reminds me star wars

[manic igloo]

Anyone who can give me a short version about this sdxl - happen to dl some models that were unusable and then noted the XL afterwards in them

[crisp owl]

New model, much like 1.5, 2.1, then there's sdxl.
Larger training base, images trained on 1024x1024 (1MP) images. Better image cohesion.
Still new, finetunes still being made, so it'll only continue to increase in quality

[manic igloo]

Aight, so I can give it a bit of time before trying it out if it's new new

[queen adder]

DreamshaperXL

DreamshaperXL + copaxTimelessSDXL (0.5 denoise)

[crisp owl]

If you scroll through some of the images here you can see some examples of what it makes now.
I use sdxl strictly now because the images are generally better when you get it right

[peak dove]

This in SDXL?

[soft zealot]

time is relative

[mellow dome]

what does rotate intent do lol

[crisp owl]

No these were MJ. He was showing that MJ has that ability now

[uneven gazelle]

Anyone having OpenCV (4.8.0) issues with deforum ?

[uncut fiber]

@mellow dome it is same seed and same all, but rotating latent space do such small differences between pictures from same batch. Probably.

checnk this 🙂 @mellow dome

i meant it is very very small difference

[spring fulcrum]

So when I am doing anything that involves skin I am getting this scratchy result... Can anyone suggest a way to reduce this effect... Here is a sample of what I am talking about... I think it has to do with the upscalers adding details to make skin look less smooth or perfect... something like that. I will also show some settings I think might be effecting it.

Any suggestions would be appreciated

[peak dove]

Upscale in stages, do a half upscale each time, two half upscales equalling a whole one.

[spring fulcrum]

do you think the upscale is what is causing that wierd skin texture?

[frigid halo]

can you use the old vae-mse-84k with sdxl?

because I feel my images have a little bit too much baked/pastel-like colors, so I would try to use vaes that adds a little bit of grey

[strong field]

I believe that one is trained on different weights so it would likely not work. In the even it didn’t throw a sizing error I would expect very unpredictable results

What do your sampler settings look like? High steps + high cfg?

I might suggest using a preconditioned latent stage with more “grey toned” latents. You can either use prompt conditioning or even revision

[frigid halo]

I use dpm 2m kerras cfg 7

preconditioned latent stage? is that a comfi ui?

[strong field]

What gui are you using?

[frigid halo]

automatic

[strong field]

Model and step count?

[strong field]

Yes was referring to comfy workflow

[frigid halo]

I was testing sd-face-swap with dreamshaper examples of gens :

as you can see a bit baked

of course it is not the character I want the likeness is not there

roop is a bit unreliable but I was searching for noses

but sdxl is a bit slow

[strong field]

Oh yeah for sure baked. Strictly talking about composition and not likeness here. Try to Raise the steps or lower cfg

Or both

[frigid halo]

ok thank you

[strong field]

See if you get better results

[ionic dragon]

whats the character called?
ebola or something?

[stone fossil]

That is a Plauge Doctor.

[ionic dragon]

use facefusion for faceswap its great

[ionic dragon]

yeah, tq

[woeful patio]

quick question, is controlnet supported yet in a1111?

[strong field]

SDXL?

[woeful patio]

yes please

[frigid halo]

ilyasivel says it should be available on august 31

[woeful patio]

ooh thats like almost here then

[frigid halo]

right before the classes

[woeful patio]

what classes? Sorry I'm not really up to date with the current developments

[frigid halo]

I mean school for me

[woeful patio]

ah yeah that sucks, school goes first 🙂

[woeful patio]

anyone know how to fix this>

[pure crystal]

I had to have the "if you can barely afford to purchase a 4090, you can't afford to run a 4090" talk with a friend today

[shy kelp]

Does SDXL have better amd support ?

[vital ermine]

you make that sound as if that issue is related to SD, which it isn't.

[shy kelp]

Hmm I am a bit clueless about the technical stuff

Is it the same but amd are the ones that need better support for AI ?

[vital ermine]

It's cool but the AMD issue is AMD not SAI's SD issue. Sad as AMD has the power but they want to keep the AI stuff for their MI lineup.

[vital ermine]

ROCm is getting better but now the issue is Python (pytorch/tensorflow) to support the latest ROCm that supports the newst consumer cards. They are dragging their heels. @shy kelp

[woeful patio]

do you have a workflow for these? they look awesome

[nimble heart]

the pytorch 2.1.0 release candidate uses rocm 5.6
https://dev-discuss.pytorch.org/t/pytorch-2-1-release-branch-cut-for-pytorch-core-is-completed/1473

[nimble heart]

the vision 2.1.0 branch hasn't been cut yet I don't think so the RC for that will take a day or two

or vision 0.16 rather

[vital ermine]

5.7 is when it all comes together, but good to know it is at least a RC out there though pytorch sucks as I am still waiting on a proper (out of dev) pytorch for 12.x/12.2 Cuda.

[nimble heart]

What makes you say 5.7 is when it all comes together?

[vital ermine]

From AMD users owners. 5.6 still has issues for some

As in Kernel panics and stuff

[crisp owl]

Select a different style

[nimble heart]

Lisa said that the XTX would be added to the "official support" list in the "rocm fall release" but whether that adds things like flash attention and all that no one knows

[vital ermine]

Yeah, no mem attention sucks

[nimble heart]

I had GPU resets in vulkan games when I first got the card but those have basically vanished on latest kernels

I've had one reset in the last like two months

[echo temple]

Hi everyone! I was looking at the reference_only CN in automatic1111, would somebody know if it is already an implementation/script for this in SDXL?

[vital ermine]

Notthe same though. I am talking about AI stuff. Train and boom possibly

[nimble heart]

Never had a proper kernel panick from this card yet

I'm not a trainer though

[vital ermine]

AMD works fine for inference

[nimble heart]

wonder if training still has the memory issues

[vital ermine]

See, I am a trainer so I never even think about just pressing a button out comes a pic side of it.

[nimble heart]

for LLMs the memory usage is identical to a 3090

[vital ermine]

mem issues it does as it needs, demands, proper memory management/attention but no dice yet

ADA Nvidia cards is having issues of our own

[nimble heart]

lol yea I see people in here constantly upgrading/downgrading drivers

[vital ermine]

I downgraded but no help tbh

I have a suspicion some of our speed issues, if not all of them, is due to pytorch/tensorflow

[nimble heart]

one thing that's nice about linux, render management, vulkan, GL, and compute are all separate drivers

[vital ermine]

we need to ditch 11.7/11.8 and move to 12.2 but can't get that working with the dev branch for SD

[nimble heart]

think nvidia still uses the one big fat 900mb driver though

[vital ermine]

I train in Linux but lately no different in speed only better memort management on Linux

yeah, it does. 1GB all in one

[nimble heart]

lame

[vital ermine]

agree

[nimble heart]

hopefully nouveau takes off

[vital ermine]

that has been aroundfor so long I thought it had died. auto turned off in Ubuntu

[nimble heart]

ya apparently it works okayishly on the 3000+ series cards

still nowhere close to 100% speed

[vital ermine]

Makes me sick that Linux is stuck on 525 or 535 for its drivers

Windows has 10+ drivers since 535

[nimble heart]

535.104.05 is the latest on my repo

[vital ermine]

ain't that the shit?

[nimble heart]

w/ cuda 12.2

[vital ermine]

Windows is 537 something 12.2

[nimble heart]

nvidia's linux drivers are basically on an lts schedule. they just port all the changes once every like 5 versions

tbf amd does the same thing with its pro drivrs

[nimble heart]

they release on linux once every quarter I believe

[vital ermine]

Well, Linus Travold despises Nvidia, as I do, but it is the devil we are stuck with for now.

[nimble heart]

the xtx would be an instant recommend if the software wasn't so jank lol

I'll have to try building amd's AIT fork at some point to see the maximum speed it can do

[vital ermine]

I almost bought it and the last second I bought the 4090 due to that fact. As a trainer I didn't need all those headaches and hope

[nimble heart]

oh you did go the 4090 finally

[vital ermine]

yes

[nimble heart]

big upgrade from the 3060 lol

[vital ermine]

I had a 1060

[nimble heart]

Oh shit

[idle zenith]

all the lora models i have in the folder isn't showing up in the web.. what could be wrong?

[nimble heart]

Play some games on your TV with it. Devil May Cry 5 @ 8k 90fps looks amazing

[vital ermine]

It was on its last legs and fans dying, needed new pads again, etc...

I am not a gamer but I am tempted

I wanna see all the candy on

[nimble heart]

steam sales are good. can get Witcher 3 for like $10 sometimes

has RT and all that

cyberpunk is cool of course but I've never seen it below $30

[vital ermine]

I really liked W3 as well as God of War 3 I think it was and RDR2

[nimble heart]

isnt the new god of war on steam too or no

[vital ermine]

RDR2 made me so pissed that it killed off my character of Arthur. Man, that hit me in the stomach.

[nimble heart]

never played, looks pretty on my friend's 2080 ti though

[vital ermine]

good game I played in 2022. Really enjoyed it

[nimble heart]

I've got enough games to play lol

[vital ermine]

Non woke was a huge plus. If a game isn't woke I will play it, and probably enjoy it so I look for indy games these days.

[nimble heart]

Still need to beat persona 5 I'm a hundred fucking hours into that thing and there's a true true final for real this time boss I haven't fought yet

[vital ermine]

Indy movies/films for the same reason

I think, I hope, when Pytorch/Torchflow is released for 12.2 officially I should see a speed boost

[nimble heart]

it's already in nightly

[vital ermine]

yes, but nothing will work with it

Everything forced me back to 11.8

[nimble heart]

idk about windows but nightly pytorch works fine on everything for me

[vital ermine]

or some LLM 11.7

[nimble heart]

even on projects which still had torch 1.12 pinned

[vital ermine]

not for me on Linux nor Windows as shit broke for training

This package demands this, that package demands that, etc...

[nimble heart]

satisfy its demands

✡️

there's no pentagram emoji so star of davis will do

[vital ermine]

I did by going back to 11.8

[nimble heart]

they all have to match so if you upgrade torch to use cuda 12 you also need to upgrade vision and triton and whatever else you use

[vital ermine]

I even had some devs tell me that is why I had errors because I was on 12.1 and they were right when I rolled back

[nimble heart]

most of them it's as simple as re-installing or if they don't have an official package you just python setup.py install

[vital ermine]

Did all that and when the dev of the program you are using tells you to roll back you roll back

I did and the issues vanished

[nimble heart]

wellp

[vital ermine]

kohya is bad about it

[nimble heart]

probably once torch moves to 12 on its mainline release other libs will move over too

[vital ermine]

yes, as I said. 🙂

[nimble heart]

it's the rocm 5.6 situation all over again lol

[vital ermine]

yep

once it rolls out from dev to official release things will work

[nimble heart]

2.1.0 ships both cuda 11.8 and 12.1

[vital ermine]

I just hope I get my proper speed then on an ada card

[nimble heart]

instead of 11.7 and 11.8

[vital ermine]

well, once they release it isn't just magically all work we need the devs and stuff to update too.

[nimble heart]

even if you're down like 10% its still the fastest card on the market I wouldn't put everything on hold waiting for 12

[vital ermine]

My hope is for the fall

You don't know?

[nimble heart]

AMD 🤝 NVIDIA

hoping everything is fixed in fall

[nimble heart]

No

[vital ermine]

Speed is as fast if not slower than a 3090 for training

[vagrant wasp]

how to run sdxl on a1111

[nimble heart]

bruv

[vital ermine]

I should be 57% faster

[nimble heart]

just put model in models folder. need the latest auto release

wonder if I could actually beat your 4090 in training rn

[vital ermine]

Did you know on XL version of kohya training 2.1 I was slower than a T4 on Colab? I switched over to Ubuntu and was the same speed. I switched to pre XL version and was around 5it/s on 2.1 vs 1.4s/it on T4.

yes

[nimble heart]

assuming I don't run out of mem lol

[vital ermine]

Still slower than my friend's 3090

I just wonder if Jensen is doing this to protect Hopper sales since it only impacts AI stuff?

Hopper sucks for gaming

[nimble heart]

which one's hopper again

[vital ermine]

the 45k chip

it is really just an accelerator but for AI look out

[nimble heart]

I mean consumer cards are already kneecapped in memory so

[vital ermine]

I wonder if there is some sort of LHR type deal happening?

[nimble heart]

amd cards don't even use gddr6x

still has a 384 bus though so maybe isn't big deal

[peak dove]

https://aituts.com/sdxl/

[nimble heart]

why reference some yoho's probably outdated guide over the official wiki https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#sd-xl

[primal vault]

Hmm, maybe you should ask the person who made that workflow?

Anyway, v3.999 - the release candidate for v4.0 - is now available on the SeargeSDXL page on CivitAI

[spring fulcrum]

That would be great....

I left you a comment of civitai about how much I like your work

[stone fossil]

Pure txt2img fun.

[spring fulcrum]

Do you have any suggestions on what I should change to reduce that skin texture effect? One effect is like a blotchiness. the other is a scratchy effect.

Here are my settings as well:

I can send you the whole image in a DM but its not somethin I can post here. ;P

[primal vault]

Start by disabling the 1x skin detailer in upscale models, then latent detailer to normal or soft, hires softness 0.1-0.2, hires detail boost 0.25-0.5, precondition strength 0.25, hires latent detailer (needs 3.999) to normal or leave it off if you stick with 3.991. You can also try to disable the secondary upscaler or swithc it to nickelback.

for realistic images, sometimes even a precondition strength up to 0.5 can improve the results, but it will change the image a bit

[spring fulcrum]

I'm going to try all of that now... I am also going to update to 3.999 after the tests and will run them again.

Thank you

[primal vault]

Keep in mind that 3.999 will not be able to load 3.991 images or json files. You have to transfer the settings manually when you upgrade. Downside of using unfinished alpha versions 🙂

[spring fulcrum]

Thats expected and fine with me... I still have a second comfy window open with 3.4 to do exactly that

I really like that you added the extra LoRA slots

[sand pier]

Cute dragon

[nimble heart]

the dragon is more photogenic than the girl

[eternal fog]

wow, just got the newest version of ComfyUI now the guy who's done the IPAdapter nodes has a fix and it performs a lot better so far.

[upbeat summit]

and 15 minutes ago t2i adapters got support for 3 channel inputs

[eternal fog]

I don't think I've used any of them, as I've been using the full ControlNets

Not sure if I know what the "3 channel inputs" are though. Not looked at that.

[primal vault]

one more, try ultrasharp as high res upscaler, that made a huge difference iirc

[spring fulcrum]

Can I DM you the image so you can drop it in and see what its doing? I tried all the different settings and still get the same.

[primal vault]

I prefer to avoid dm for helping, because nobody else can see the solutions

[spring fulcrum]

You can come back here for the help I just cant put the image in this chat

[crisp owl]

Geez I was just setting up an ipadapter flow and ran into that error, must have literally been 21 minutes ago, and the update was 20 minutes ago, because I checked for an update after the fail lol

blink and you miss an update lol

[primal vault]

I don't need to see the image to help with the issue 🙂

[spring fulcrum]

thats true

[rustic surge]

help me please!!

Help me, how to solve this error

[spring fulcrum]

I just upgraded to 3.999... Running the tests again

[primal vault]

no detail processor and ultrasharp for high res upscaler, try those first before tweaking other settings

[zinc cargo]

looks so cool

[eternal fog]

You need to get the Dev version of the IPAdapter.

A Comfy Update broke it, he's updated it, but because it can break older versions he's put it in dev for now.

Go into the custom nodes folder for IPAdapter in a command prompt and use git checkout dev and then git pull

Just remember to keep an eye on updates so you can switch back to stable once it's pushed there.

lmao the text it's put on this person, I did not tell it to do that

[glad grove]

i support the suck initiative

[eternal fog]

Nice images. It's cool what you can get when you put stuff like this into IPAdapters

[upbeat summit]

I just saw the additions on the SDXL Prompt Styler node. Very cool 🙂

[soft zealot]

ah I do need to update really (although I'm ok with the way I have my setup)

[upbeat summit]

Just saw you mentioned in the repo README.md and wanted to say thank you for enhancing our toolkit 😉

[soft zealot]

what you may not have noticed (going off on a tangerine) is my lighter weight workflow variant, take a look at the last image I posted ;o)

[upbeat summit]

oh - just loaded it up. that is new. looks really nice and tidy.

[soft zealot]

its a lighterweigh simpler implemetation

[upbeat summit]

yeah. I really like your advanced workflow as well 🙂 but this looks very nice too. you have a chatgpt node included - cool!

[soft zealot]

thats in my main one as well now

[hardy cipher]

testing out my doppelganger machine

very accurate

[upbeat summit]

You're working on crazy machines again, Mr. mad scientist picturesonpictures?

[hardy cipher]

using the canny node I put together. I need to submit the nodes to the manager I think. people might actually want to use a couple of them

[upbeat summit]

that sounds cool 🙂

[hardy cipher]

not sure how long the regular canny edge finding node takes for most people, but the one I made is about 50x faster on my setup

I mean, the other one seems like it's more complex, lol

[eternal fog]

Isn't the Canny Edge pretty much instant?

[hardy cipher]

which is probably why it takes 50 times longer

I mean, the main one that I compared it too, "canny," isn't pretty much instant for me

but maybe for some people

[eternal fog]

I just double checked and yeah it's instant

[hardy cipher]

is it?

which one?

[eternal fog]

Are you meaning the generation of the actual image, or just getting the Canny lines out of an image?

[hardy cipher]

yes, the canny lines. there are multiple. I didn't compare all of them. just the default one from what appears to be the default controlnet nodes

either way, this one has different settings so you can adjust things differently

[eternal fog]

The one I have is just called "Canny Edge Pre-Processor"

[hardy cipher]

I'll see how it compares. the one I compared to took like 4 or 5 seconds. but my computer isn't exactly top notch in comparison to what a lot of people have here

might not be breaking new ground with speeds, I don't know. but I do believe the settings are a bit different

[eternal fog]

That's the one I have

Hands plz 😢

[hardy cipher]

cool. I'll see if they are comparable. it's not like I did anything incredibly groundbreaking. just found libraries that did what I need

have it do a quick danbooru interrogation of the image, then add some default words to the end and use that as a prompt for the canny controlnet

doesn't necessarily make the image better, lol. but results can be interesting

divine beauty

[upbeat summit]

this is the only style I will tolerate and accept these deformations 😄

[hardy cipher]

I sent some of them to my mom and she told me how good they were. 😭 . she didn't realize I was making them ironically and surely had to think they were goofy looking

[upbeat summit]

I think you should take this is a positive. It's your mom and maybe she just has great taste in modern art 🙂

[hardy cipher]

this is one I sent her

[glad grove]

my beloved 😳

[hardy cipher]

tbf, it is pretty classy

[upbeat summit]

To be honest it's not as bad / funny / cool as raw TerribleFanArt. It has nice shading and lighting and looks overall like a high quality caricature.

So your mom is right!

[hardy cipher]

Thiccneck Seagal

[upbeat summit]

he looks like a thumb

[hardy cipher]

remember when he teamed up with daddy long neck?

https://www.youtube.com/watch?v=kH00Kms0eC0

this is why the internet was created

just noticed they also included missy long tongue

[upbeat summit]

I haven't seen this before, so I'm now smarter because I've learned new internet lore

[hardy cipher]

yes, daddy long neck is a bit of a rabbithole to go down. gucci berry is the real gem

[visual glade]

the canny node in base was using the CPU, I updated it so it uses the GPU so now it's fast

[gray loom]

/d

[atomic ice]

can I question you?

I need you to be questioned

I have an amazing dataset of 512x512 images

can I use this dataset for an SDXL LoRA?

or do I need to cry and make a new dataset?

[eternal fog]

You can use it, but the images that it produces might not be brilliant quality

[atomic ice]

So I need to cry?

[glad grove]

upscale the images with a good upscaler then u can use them

[atomic ice]

also, can I train just like I did with my amazing 1.5 models with the captions in the file with the same name?

[atomic ice]

WOW!!! YOU ARE 131314144IQ!!! THANKS!!!!

HOW DID I NOT THINK ABOUT IT?!?!?!

I'll do it!!!!

thank you x1000000

[glad grove]

https://tenor.com/view/steven-seagal-stare-look-serious-intense-gif-5002551

[indigo carbon]

just looked into it; this is what SDXL makes on the first try

[void skiff]

I use SDXL with Automatic1111 and I can get some good images, but I see that the SDXL bot here generate better images that follow the prompt better and give me larger range of variations. So my question is what I should do if I want to get similar result as I get here with my Automatic1111 when working locally, or should I try another interface.

[visual glade]

the SDXL bot here uses ComfyUI

[native knot]

ComfyUI>A1111

It's got a ton more flexibility.

[upbeat summit]

behind the scenes information

[void skiff]

Ty, I will look into ComfyUI I am just a bit afraid for spaghettification, but I may learn 🙂

[native knot]

Start slow. Build a very basic workflow. Spend time understanding what things do, then add to the workflow to try things. Don't be afraid to break anything because when you do, you simply go back to an image you made that worked and drag it into the interface...then boom! You're right back to the functioning workflow.

[indigo carbon]

UPDATE ON THIS: turns out it was a troll. the images that were shown off were by users and were made by top gens of SDXL LOL

so yeah, the only instance of MJ being better than SDXL were actually SDXL making good than usual gens

[eternal fog]

lmao, I think this might be bullying, it's written Ginger on her uniform

[void skiff]

First image made here, the other one in A111, the magical brew was "a image of a pink funny character with a ponytail and pink nose BREAK wearing a yellow dungaree suit"

[glad grove]

BREAK nose BREAK legs BREAK knees

[void skiff]

funny man BREAK dance.

[uncut fiber]

TB woke up out of control!

[uncut fiber]

Inpainting or normaly generated?

[glad grove]

i dont remember this version of the carebears

[vital ermine]

I can't use inpainting in comfy

[uncut fiber]

ok just asking because small difference

[vital ermine]

Yeah I changed from red hair to reddish brown hair in the prompt

[uncut fiber]

oh i forgot. I got same seed same prompt now, and as well small difference

[trim hollow]

/production du lait

[vital ermine]

oh, man

[vital ermine]

Into this for me

[heady vale]

into this?

Im still waking up here

[vital ermine]

@heady valeor this

[heady vale]

not much of the original image there

[vital ermine]

which is fantastic.

I can make it more but I am using it for the colours

subtle stuff

[hoary saddle]

did you ever get a chance to try adding cnet into your AIT workflow?

[indigo carbon]

I did actually manage to compile modules for SDXL CNET, but due to ComfyUI limitations; I wasn't able to implement it. but it did work with a demo script

so after ComfyUI will be more compatible; it will come

[void skiff]

My first ComfyUI render that was not a bottle

[indigo carbon]

also, it seems impossible to compile CLIP_vision as of now.. might potentially be possible with future versions of Facebook's AIT or even with the current FX2AIT. will have to look into it

[void skiff]

But when I saw ComfyUI I thought this look complex but I understand it, but then my fear came, what if I want to do those useful X/Y/Z charts, how to do Img2Img and all fun things with Control net, AHHHHH!... okay, take it easy now, remember, start with baby steps.

[hardy cipher]

https://tenor.com/view/oh-ok-gary-payton-what-well-ok-then-alright-gif-15987992

[vital ermine]

Everyone tells me XL can't do Anime. Otay.

[hardy cipher]

damn

that's pretty top notch

[upbeat summit]

thank you! uses a new beta of NightVision XL by @delicate kelp. Here is the prompt:

dynamic pose, neural net dreaming wood elf in a fight looking at you, unreal engine 5 render, a detailed matte photo, photorealism, crystal clear, (high contrast), (gorgeous photo:1.3) of (ultra detailed:1.3), 8k, masterpiece, (ultra detailed:1.1), 16k textures, cinematic look, intricate, epic, UHD, DSLR, raw photo

• your favorite negative prompt

this is a new prompt build I'm working on - work in progress.

[hardy cipher]

neat. I think I need to update my models. they're so ancient now

[steel mist]

The one thing that SDXL has really blown me away with is lighting

The like, lens aberations on that elf are amazing

[steel mist]

I need the opposite—remove aberations from my actual photos

[vital ermine]

@upbeat summit that was your prompt with my new lora I am about to release

wow, hmmm

[upbeat summit]

yes, but to be fully transparent here - the image runs through a small post-processing pipeline I'm using in ComfyUI. raw output -> upscale -> a bit of film grain -> chromatic aberration

here is the raw output image from the model before being post processed:

[steel mist]

ahah well that explains it

the lighting though... still superb

[steel mist]

I actually hate CA in generated images and games 🙃

[eternal fog]

Accidentally managed a sort of Real + Anime look

[steel mist]

Like, I spend ungodly amounts of money to avoid it in camera gear, and then these games end up spending GPU cycles to duplicate it

[upbeat summit]

yeah. I just add a bit of film grain to give it more subtle perceivable details and the chromatic aberration to emulate a camera sensor look. but I only do this if I want to go full cinematic with the image 😉

[steel mist]

Yeah big fan of noise. Honestly I want to spend a little more time getting photoshop integrated with some of these workflows

something like A111 in a photoshop plugin would be really helpful

though I finally reinstalled A1111 and holy heck it uses VRAM like there's no tomorrow

[upbeat summit]

I grabbed a couple of very high quality analog film grain frames from my video assets I use for other projects and it works quite good. I just save all intermediate versions, so I'm never locked in

[vital ermine]

I died. I prompted Dr Fauci

[hardy cipher]

did you put skeleton robot ghoul as well?

[vital ermine]

no

XL said this is him

maybe in his true form

[hardy cipher]

wow, well that's that I guess. XL knows best

[vital ermine]

Aye

[upbeat summit]

that's my very simple post-processing pipeline in ComfyUI and I enable the nodes depending on what I want to do

[vital ermine]

I need one of those crazy ultra ultra wide monitors for comfyui

[uncut fiber]

do you mean for 1536x640? @vital ermine

[vital ermine]

one of those weird 30:1 type ones I have seen. does 4k but stretches two desktops

[turbid garnet]

portrait of a model with raindrops in the foreground, wet look, slight smile, hyper-detailed skin texture illuminated by neon light, white dress, long hair, fickle face, perfect anatomy, wet-to-wet background technique by Misilbu. . hyper detailed 8k painting, 8k concept art portrait by Greg Rutkowski Artgerm WLOP Alphonse Beeple Caravaggio, muted colors, watercolor style, bokeh, f1.0 lens

Made with SDXL from Monster API - https://developer.monsterapi.ai/reference/post_generate-sdxl-base

[uncut fiber]

i got negative prompt, which should not allow this.

[slender coral]

Anyone know the tech behind pika labs? https://www.pika.art/

[hardy cipher]

it gave her long face for some reason

[eternal fog]

It do be like that sometimes

[hardy cipher]

lol

[hardy cipher]

could probably have fine tuned things a bit better. but that thing took a little while on my poor little laptop

[uncut fiber]

Hotel rooms

[eternal fog]

nooooooooooooooooooo

[hardy cipher]

I still remember a nightmare from when I was about 5 that resembled that

[vital ermine]

The money shot

[hardy cipher]

a couple bugs on the wall, they started growing. kept growing. grew larger than me. not what I wanted to deal with at 5. or now

[glad grove]

average hotel room in australia

[uncut fiber]

nightmares are realy bad

[vital ermine]

Dunno what that is. Run if seen. The allure though. Damn.

[hardy cipher]

perfect

[vital ermine]

https://civitai.com/models/137273?modelVersionId=151473

[upbeat summit]

congrats to the release! I will check it out in the morning!

[vital ermine]

Thanks.

I am loling over civit using Goth Jesus as he was not the first image

Civit hiccups a lot as it demanded I upload and I already had so I redid it and had two of everything. :/

[upbeat summit]

their servers are very busy

[vital ermine]

overloaded is right

[upbeat summit]

there's only so much you can do with lots of traffic and users. they probably have to rebuild or refactor some of their backend code to make this faster.

[vital ermine]

Yep, and may need to increase bandwidth too which requires paying more for bigger servers. The plight MMOs go through

[uncut fiber]

Weird Thunderbirds 🙂

[glad grove]

look like one of those weird lookin enemies from tokusatsu genre

[uncut fiber]

yes there are some TB like this, probably some different version

[steel mist]

sorry if this was already brought up, but did an update break the advanced nodes? Don't see anything in manager that needs update, but getting this error

default ksampler is fine

[sterile yoke]

i'm shopping for a new computer; what are some things to look out for if i want to run stable diffusion xl on my computer?

[glad grove]

nvidia card,rtx and lots of vram

[sterile yoke]

ok
when it says like "12 gb memory" is that the vram

[steel mist]

Yeah

[sterile yoke]

how much is lots

[steel mist]

if you're buying a consumer GPU you'll also game on? The answer is 24 gig

[sterile yoke]

gotcha

[steel mist]

SDXL needs at least 8, but my 2080 8G startting running out when doing more advanced stuff

[glad grove]

if u going for a cheap one just buy a 3060 12gb edition

[sterile yoke]

https://www.costco.com/lenovo-slim-7-pro-x-14.5"-touchscreen-laptop---amd-ryzen-9-6900hs---geforce-rtx-3050---120hz-3072-x-1920-display---windows-11.product.100976621.html
how does this computer look

[glad grove]

laptop no good

[sterile yoke]

?

[glad grove]

plus that one only has 4gb vram,cant run sdxl on that

[steel mist]

laptop thermals tend to suck and the GPUs aren't as good as a desktop equivilant of the same number

[sterile yoke]

well there's no chance i'm buying something that isn't a laptop

[sterile yoke]

where do you see that?

[sterile yoke]

oh gotcha

[steel mist]

Thought you were talking desktop cards; usually the vram is pretty obvious on those

anyways I know basically zip about windows laptops

[sterile yoke]

i currently have a windows laptop and would like to stick to something familiar
plus i like putting the laptop on my lap
plus i don't really have the space for a desktop
plus i like to be out in the living room sometimes

[glad grove]

if u want a laptop it would probably be the "Legion Pro 7i " that one has a 4090 with 16gb of memory,price is $3k

[steel mist]

lol how do you even power that

[glad grove]

battery probably lasts like 30mins

[steel mist]

the TDP on a 4090 is like, 4x the USBC power draw?

I guess it's a mobile part

[sterile yoke]

pffff
the absolute most i'm spending is $1100

[steel mist]

I would look into cloud services then (eg, google collab)

[sterile yoke]

i have

[glad grove]

or u could also go for the "ASUS ROG Strix Scar 18" price its $4k

[sterile yoke]

google colab is nice if you can get it to work (big if)

[glad grove]

but if u only have $1k thats not enough for a good new laptop

[sterile yoke]

does it have to be new

[glad grove]

no u could get an used one with a 3060 12gb

that one would probably be your best option

[sterile yoke]

lemme see what i can find

[visual glade]

make sure you don't apply the control lora to a model that doesn't support it

[vital ermine]

Then I would suggestthis is not for you OR, if you are hellbent on only a Laptop, there are some 5-10k Alienware out there that would work.

[sterile yoke]

5 to 10k dollars???

[vital ermine]

1100 will not get you anywhere near what you need for this on a laptop. Maybe, but I highly doubt it for XL.

[sterile yoke]

feh

[hardy cipher]

alright, nodes submitted. watch them get denied, lol

[sterile yoke]

ok i guess i'm not buying a new laptop for a while then

[steel mist]

That was not it, but it was a similarly boneheaded move 🤦‍♂️

(wrong clip)

[hardy cipher]

I thought my nodes were broke beyond repair because I didn't realize I had to actually remove and re-add them to the interface even after refreshing

went on like that for quite a while too

[sterile yoke]

speaking of google colab
does any of you know a good sdxl google colab that you can use with the free tier

[hardy cipher]

I'd imagine comfy would be as good as any

probably less vram usage than a1111, although I haven't been keeping up on that

[vital ermine]

it is less vram

[sterile yoke]

the problem with comfy is that it doesn't seem to have inpainting

[hardy cipher]

it surely does, right?

I haven't actually looked tbh

[vital ermine]

it does just a pita to work in. Not nearly as bad as its outpainting, but inpainting mask it is best to use Gimp or PS for that

No zoom in on the canvas unless someone made a new node for that

[sterile yoke]

is there one that isn't comfy ui?

[hardy cipher]

there are several that aren't comfyui

but you might have issues with the free colab

[vital ermine]

damn it, I had to unpublish my lora as it had it listed as 1.5

I can't find a way to change that

[sterile yoke]

i'm trying one by camenduru but i can't get it to work

[vital ermine]

there we go, fixed

[hardy cipher]

you can resubmit right? I know that's not what you want

[vital ermine]

Nah, it was in the drop down. Civit is weird as it should be in the model info but they separate version info and model info. All fixed now

[sterile yoke]

can you send me some links to those in dms or something

[hardy cipher]

civitai is a bit of a train wreck. I mean, better than I'm doing in the website department, but it's a hassle

[vital ermine]

Yeah. I stayed away from civit but HF had zero traction, and was never set up for this more for researchers and it had its own issues. At anytime they could come do a clean sweep too I was told.

Seems that if I want traction I need to go back to 1.5. Downloads of 2.1 is the worst, XL second worse and 1.5 is something like 4-10 to 1 on civit

[steel mist]

I would just like a filter that shows me NSFW stuff but like, not explicit furries?

[glad grove]

i agree they need to add a furry tag

[sterile yoke]

they do

[vital ermine]

is there an anti furry tag?

[glad grove]

i think thats just normal porn

[vital ermine]

well, a lot of NSFW I don't consider to be NSFW but it does so it is all full blown worse than porn hub or Sistine chapel.

[glad grove]

when searchin for nsfw,just dont scroll down to the bottom

(real )

[hardy cipher]

I'd suggest trying stable swarm. it uses comfy as the backend, but it's pretty much plug and play in the main window. then if you want to get ambitious you can switch over to the nodes. https://colab.research.google.com/github/Stability-AI/StableSwarmUI/blob/master/colab/colab-notebook.ipynb

[sterile yoke]

thank you so much

[hardy cipher]

it's still in it's early stages so not super refined or flashy. but does the job, and less vram usage than a1111

[sterile yoke]

thanks

[vital ermine]

You mean when not searching? I have NSFW filter on cause the first time I went all I had were pricks and bushes and other things all over my screen. LUCKILY, I was not at work.

[glad grove]

yea the bottom is where the very specific and low quality content lies,you will see all kind of crazy stuff there

[steel mist]

content is coming out so fast though I've kept the filter set at newest first

[indigo carbon]

@visual glade were you able to use AIT on CLIP_vision? I tried to compile using normal CLIP compiler and even FX2AIT; it seemed like it wasn't possible. any idea?

[visual glade]

does CLIP actually benefit from AIT?

[indigo carbon]

it's required.. if you don't compile for CLIP as well as UNET, it won't be able to do any sampling- this is why I attempted to compile for CLIP_vision and CNET

[hardy cipher]

picturesonpictures — Today at 10:30 PM
I sometimes wish there was a way to see who the degenerates are that create some of those models and loras. are they just regular dudes that act all normal in their everyday life, then make nsfl furry scrum loras in their off time?

[visual glade]

it works if you just compile the unet though

[indigo carbon]

it depends on how you compiled UNET. if you compiled JUST the UNET it won't work without the CLIP being compiled into the module. it's just that when importing the pipelines before compiling a checkpoint with specified architecture, it compiles the entire checkpoint, including CLIP. but that's not the case with CLIP_vision so it needs a script of its own to compile

[vital ermine]

One can only imagine.

[indigo carbon]

also FX2AIT should be independent and able to compile everything; but it still doesn't understand CLIP_vision

[hardy cipher]

it actually had me really turned off to the whole text to image ai community. until i found places like this it seemed like 80-90 percent of everything was extreme waifu furry creeper stuff. not that there's anything wrong with anime stuff, but ugh. the levels they take it to

[vital ermine]

Oh, 90% of my discords I deleted were full of anime, waifu, furry trainers. Didn't care for realism and were all 1.5 so I split them after a bit realizing I am the odd man out in them.

[indigo carbon]

@visual glade if you try to use CLIP_vision when using AIT in the K_sampler; it's unable to do any sampling. it spits errors that don't fit into discord. this means that CLIP_vision requires to be compiled for it to work with AIT

[hardy cipher]

what's a bummer is some of those models are legitimately good other than the nude anime stuff. but you really can't escape it. I'd put nsfw, nudity, etc, in the negative and it'd still shine through

just too overwhelming for the model to suppress it

[vital ermine]

In SHARK it takes a good while to train the model but in AIT it is almost instatenous so I wonder if it really compiled anything?

[vital ermine]

Due to how strong some are. Corset works on some but most times not even then.

[indigo carbon]

AIT doesn't include weights in the modules, it just gives it a way to load a checkpoint using AIT- this is why AIT isn't checkpoint specific

[glad grove]

because thats what most of ppl who draw anime style create

[indigo carbon]

there is also another thing AIT is capable of doing called "FX2AIT", I'm not sure if it works- but it's supposed to be able to compile ANY AI model. but I never got it to work

[vital ermine]

That says something for sure. Anime has become a sub culture it seems. :/

[glad grove]

yea but in the 80s and 90s was more "independent" now that its becomin more mainstream its gettin more censorship

[vital ermine]

Japanimation renamed Anime for woke reasons and look at it now.

[glad grove]

now you rarely see one of those gritty gore vhs ovas from the 80s 😔

[vital ermine]

Hentai was were the dark was back in those days.

[glad grove]

even the porn ovas had some crazy edgy gore reminds me to those explotation films from the 70s and 80s

but if u want non porn anime models you could try with some 2.1 models i have seen some that are decent enough

[hardy cipher]

I like the artwork. it's great. but can't get with the weird overly sexualized stuff

[vital ermine]

I loved Japanimation but Anime turned bad so now I just don't really dig any of it. If I want porn I watch real porn not cartoon shit.

I wonder if non sexual Anime still exists or is that not a thing now?

for films I mean

[hardy cipher]

maybe on netflix

[vital ermine]

no thanks

[glad grove]

it exists you could watch something like "makia" a 2018 movie or one of Makoto Shinkai movies

[sterile yoke]

is there an image to image or inpainting option for this?

[vital ermine]

I know that name. Where do I know that name from? hmmmm

[hardy cipher]

tbh I haven't used it that much as I'm running with the comfyui approach. but there are definitely inpainting options for comfy https://comfyanonymous.github.io/ComfyUI_examples/inpaint/

[hardy cipher]

not exactly sure how it'd work with inpainting, but in general you can load the workflow into the node window and then run it with the interface in the primary window

[vital ermine]

There we go

WOW

I turned on my lora

Calling more lora of the above

[sterile yoke]

ok i ran it and was successful, then i ran it again and it gave me this

thank you for helping me by the way

[hardy cipher]

hmm. not sure, man. do you have it on your google drive?

no problem. I try to help people when I'm able. I had to struggle through a lot of the stuff myself

[sterile yoke]

i have these folders/files in my stableswarmui folder

i don't know where the launch linux thing would be

[hardy cipher]

did it actually load for you? were you able to use it?

[sterile yoke]

yeah but then oh wait

ok i had to delete it and reinstall it because at first i picked stability api

but it got mad cause i didn't have enough credits

so i deleted it and reinstalled it with the local option or whatever

[sterile yoke]

and then it pulled up a janky website not loaded version that didn't work, then when i refreshed the cell in google colab it did this

[glad grove]

pain of free colab 😔

[hardy cipher]

https://tenor.com/view/diplein-rubens-hide-the-pain-museum-masterpiece-gif-18417428

[vital ermine]

Where do I go to get controlnet for XL?

grabbed a workflow and I need it

[glad fulcrum]

https://huggingface.co/stabilityai/control-lora

[vital ermine]

is that what it wants?

[glad fulcrum]

yes

[vital ermine]

oh, cool. thanks

[sterile yoke]

ok i think i got it to work

[indigo carbon]

hey, would it be cool if I were to make an image-mixing version of my AIT workflow? I can, figured it out =]

[hardy cipher]

I'd say yes

[opal yacht]

I'm on the same boat; I generate everything on a 11-year-old computer. I've been researching a lot and the easiest way to generate SDXL images using colab (and the best I've found so far) is to use the modified version of Fooocus, by Moonride, it has lots of features, much more than the original Fooocus, and some that are really nice, like native SAG to balance the plastic vs. overcooked view, it has Revision + Img2img , you can change samplers, etc. It uses comfy in the background, so it has all the great memory optmization, so it runs on free colab, but I've decided to pay because I generate too much stuff. And you can easily just paste a lot of models directly from Civitai (yeah no need to download/upload anything to Googledrive), and it loads really quickly for around 6 models + Loras. If you load more than 10 models, it takes a little while, but it still works fine.

Click on the play button (yeah, only one play button and nothing else) and enjoy a "it just works" approach. Sometimes, if the dev is messing with it, it can give errors, but it will work again after a few hours.

The only problem I have with this is the OOM errors from time to time if you use lots of Loras, and sometimes it slows down or the Gradio interface gives errors, but this is not Fooocus fault; it is a colab + gradio problem.

For comfy, I found a colab that already has the manager installed, but it gives a lot of weird errors on several workflows. The official comfy colab should have the manager installed already for us folks that only use colabs.

Link for the Fooocus colab by Moonride:

https://colab.research.google.com/github/MoonRide303/Fooocus-MRE/blob/moonride-main/colab.ipynb and https://github.com/MoonRide303/Fooocus-MRE

[glad grove]

ima try the mre fooocus on 16gb to see if its better optimized than the normal fooocus

[indigo carbon]

yo?

I think I did it

[indigo carbon]

does this look refined enough for a release? it's very good

[glad grove]

it looks clean af

[indigo carbon]

still uses AIT, so same speed as normal AIT workflow

I'm actually so proud of myself rn

currently testing batch size, will update =]

[opal yacht]

👍 , It probably won't be because of all the extra features. I've suggested that Moonride implement IPAdapter; it would make his version really powerful, and some form of upscale is really needed. I was surprised to see that Fooocus colab is even easier than Lastben's colab/Auto and more powerful than Auto for what I need.

[indigo carbon]

can anyone give me more stuffs to try?

[crisp owl]

a watch and a galaxy

[vital ermine]

How do I turn this off in comfy?

[indigo carbon]

will do next, let me make 1 of each real quick

[indigo carbon]

it got confused with the clock, but I think if I change the weights it will get better

[crisp owl]

hah interesting it turned to a coin

[indigo carbon]

ok, here it did a better job than MJ could dream of doing

[crisp owl]

ah nice, that's a good merge there

[meager canopy]

https://tenor.com/view/eh-meh-jim-jim-halpert-the-office-gif-25317293

[crisp owl]

ask and ye shall receive

[meager canopy]

Looks interesting! Will you be sharing the workflow...please?

[indigo carbon]

still working on it. I think I'm getting closer to perfect

[indigo carbon]

does this look clean?

[meager canopy]

https://tenor.com/view/unbothered-waiting-patiently-whatever-ill-wait-gif-18144151

[meager canopy]

Looks good to me.

It's ComfyUI, wasn't designed to look clean 🤣

[indigo carbon]

DUDE THIS IS WHAT I SAW IN MY HEAD

it placed the dog in the chair without me asking it to =]

[soft zealot]

This is about as clean looking a daily interface as I can manage 🙂

[vital ermine]

Pretty hot

[hardy cipher]

mine could be clean with a bit of effort

[peak dove]

Werf!

ComfyUI, Sytan's w/flow, "We're at the Bowl!!!"

[peak dove]

My Kittnz!

[indigo carbon]

?

[peak dove]

"Did your excellent dog on a m/cycle never go "WERF!!!" ?! 🙂

[indigo carbon]

I didn't prompt anything, I'm working on a workflow that mixes 2 images while using the AIT workflow I made

[peak dove]

Great, is an AIT Like Merge in MJ?

I must see if I can get some AITs at Civit. It looks fun 🙂

[indigo carbon]

no. a few days ago I released a workflow that doubles performance using AIT. it's a new optimization by facebook

[peak dove]

Vamp!

[peak dove]

OK, but it does look like an SD photomerge! Glad to see fb making a contribution! 🙂

Perhaps I should have said a more creative and extended photomerge.

[indigo carbon]

you didn't understand. the original workflow I made is just a good txt2img that works twice as fast as ordinary ones. I'm working on an image mixing version.

[sand pier]

More then that on certain gpus

[indigo carbon]

this is the prototype, it starts to come together. If I'll release this it will probably the first workflow that just mixes 2 images without text input

[sand pier]

I benchmarked it with 3x the performance if it's that ait one

[soft zealot]

so using the ReVision technique then?

[indigo carbon]

no. ReVision doesn't work with AIT. I worked on something that makes a conditioning that fuses 2 input images and turns it into a conditioning that didn't change the structure of the AIT loaded model. this way I can blend 2 images, while keepinng the massive speed boost that AIT does.

[soft zealot]

so ReVision like then?

[indigo carbon]

pretty much.. but isn't revision just mixing concepts instead of subjects?

it's more similar to MJ's image blend thingy than anything

[soft zealot]

revision takes one or more iimages then uses clip Vision to Encode them in the same way a normal conditioner does for text then by weighting etc you create a new image based on the interpratation of the roginal input images (it uses a blip like (but not acactly the same) mehtod to interrogate the images before encoding/conditioning them

[sand pier]

Did ait fix the controlnet bug?

[indigo carbon]

if I'll release this, won't this be the first workflow that not only uses AIT, but also mixes images?

[meager canopy]

If you don't hurry up and release it, it won't be the first! 🤣

[soft zealot]

well If you hadn't said otherwise this looks & feels like Revision.................

[meager canopy]

Is this the first banana launcher?

[indigo carbon]

yo? revision doesn't mix images.. if I'd do this with revision, it will just make a dog in cyberpunk, not a dog riding a motorcycle in cyberpunk

[heady vale]

that would be a reverse thrust banana launcher 😜

[crisp owl]

I'll take two please and thank you

[soft zealot]

my apologies, just looks veryt ReVision like to me

Isnt the workflow almost this with a few AIT modules tackked on??

[indigo carbon]

if you would use AIT in that workflow it will give you errors that don't fit in discord

[meager canopy]

Doesn't really matter what it is, as long as it works as intended 🙂

[soft zealot]

never said it would work if you just checked them in, I said it looks very similar 🙂

[indigo carbon]

guess I'll keep it closed for now then, will get more confident with it

[meager canopy]

Let us play with it 😦

[indigo carbon]

maybe, testing more stuff for now to make it even better

[peak dove]

Its' great that you're doing this stuff! I am excited at the prospect of a working model! 🙂

[indigo carbon]

my goal isn't to make a workflow that works inefficiently, I aim to make the most efficient and good workflow that does those things nicely while maintaining precision. I made a txt2img workflow that achieves that by using AIT- but now I'm working on a version that blends images

[hardy cipher]

personally I prefer to make extremely inefficient workflows

[indigo carbon]

I made this, it makes images like demonstrated in the showcase in under 14 seconds(ONLY ON 3000 SERIES AND ABOVE):
https://civitai.com/models/133818/tdg8uu-aitworkflow

[soft zealot]

FYI might be helppful to addd the comment about 30 series and above on the CIvitai description, currently it doesnt mentuion that.

May save you some stressy messages ;o)

[hardy cipher]

nah, you make some top notch realistic stuff. but my issue I haven't had good luck with ait. what's the cutoff for it benefitting video cards? is my 3060 too primitive? or am I doing something wrong somehow?

[indigo carbon]

you first need to follow the instructions provided in the description, they explain it pretty well. also make sure to have torch2.0.1+cu118 to benefit as much as possible, also the latest cuDNN version available.

[hardy cipher]

ahh. you know, I wasn't exactly thorough with all that. I'll have to revisit it and see what I can do

I don't know if I even noticed those guidelines

[indigo carbon]

it should also be mentioned on Comfy's official repo. it's crucial to have the latest firmware to use as much potential as possible that GPU has

[hardy cipher]

I don't think mine is outdated, but I'd have to confirm. don't know off the top of my head what I have installed

I indeed have torch2.0.1+cu118

but not sure about cuDNN

[indigo carbon]

cuDNN is also very important.. idk why Nvidia only made it accessible to dev accounts. we should check in with them to make it easier to get your hands on

it slightly improves inference, but not AS important as CUDA11.8

[hardy cipher]

hmm. I'll look into it. I don't recall having a dev account

[indigo carbon]

Nvidia are such assholes, every new gen they double the price and half VRAM bit capacity, only thing better with 4000 series is the newer CUDA cores

when it comes to raw hardware, AMD is insanely better than Nvidia, it's so upsetting that if AMD had software as good as Nvidia they would have destroy them

[hardy cipher]

yeah. that's a bit perplexing to me. not sure why they haven't kept up on the software side. I guess it all comes down to $$$ and nvidia has a lot of it. but it's not like amd is some little startup

[signal harbor]

Is there a good idea for automaticly variating words inside a promt in confy UI?
FE: blonde - brunette - ginger ?
Or s.th. like wildcards? -As we know from A1111?

[vital ermine]

Yep, 100%

[hardy cipher]

lots of prompting tools if you look through the nodes in the manager. not sure which one best suits what you want

well isn't that nice https://civitai.com/models/135007/adolf-hitler

[signal harbor]

Thanks!

[azure oxide]

the dynamic prompt syntax works in comfy by default. {blonde|brunette|ginger}

will randomly pick from 1 of the 3

[hardy cipher]

well there you go

[signal harbor]

I tryed that out, but only got brunette, which was the last of three words...Will try it again.

[hardy cipher]

for some reason I had it in my head that it wasn't an option by default, but I may be wrong about that

[vital ermine]

@high skiff Here is my 4090 at idle temps

[azure oxide]

its in the notes section if you scroll down https://github.com/comfyanonymous/ComfyUI

[high skiff]

yeah, looks like you don't have the issues either

IDK why I do have the issue

[azure oxide]

its your 2nd hand card 😏

[high skiff]

it happens with all my cards

new or used

[hardy cipher]

indeed, you are correct. I'm not much of a wildcarder with prompts so haven't used that much myself

trying to think of some weird stuff to do with a node though

as far as prompts are concerned

[rain crown]

Do you guys recommend ComfyUI or A1111 currently for SDXL?

[indigo carbon]

its just so sad.. reminds me of that character from DBZ that had the potential to be the strongest, but because they never trained, they were just decent

[hardy cipher]

https://tenor.com/view/oh-my-gif-25288613

[high skiff]

Comfy. Its faster, more capable, more diverse, uses less resources, and over all is just a better SDXL experience

[indigo carbon]

if AMD would have something as good as CUDA, they would DESTROY Nvidia like the assholes they are. they would also benefit from it market wise

[rain crown]

Bet, thanks my guy

[indigo carbon]

like, when it comes to raw specs, the 7900xtx is the strongest consumer GPU

[high skiff]

that is not true

[indigo carbon]

but the software holds it back soo much

[high skiff]

with previous gen, AMD had the raw perf lead, this gen, the 4090 is considerably more powerful in raw perf

[indigo carbon]

I wasn't talking about performance. I was talking about VRAM, core amount, the physical things

[hardy cipher]

if you want to start off with preset workflows there are a bunch of places to download them. and by default any image made with comfy has the workflow embedded in it so you can just drag and drop an image into your interface and it'll load whatever workflow made it

[high skiff]

in raw Teraflop performance, the 7900XTX is 61Tflops, whereas the 4090 is 83 Tflops

[rain crown]

Yeah 4090 is better even without DLSS 3

[high skiff]

thats the rawest form of performance measuring for a GPU

and if you get to RT-TFlops, the 4090 has 191 RT-Tflops

[high skiff]

I have no idea where you are pulling that

[signal harbor]

It works now, I mad a mistake using a false type inside the brackets...

[rain crown]

Yeah I have it fully set up, just wondering since I'm a lot more used to A1111 workflow, gonna just have to learn the node workflow as I go i guess

[indigo carbon]

192ROPS vs 176ROPS

plus, AMD is much more generous with VRAM. I'd grab radeon any day over Geforce if the software stack was as good

[hardy cipher]

it's not as intimidating as people make it out to be. lots of workflow examples here https://github.com/comfyanonymous/ComfyUI_examples

if you want to feel out how things work

very helpful

# Q what's the difference between a normal image and a latent image?
# A a normal image is stored as a vector of numbers that represent the color of each pixel
# A a latent image is stored as a vector of numbers that represent the color of each pixel

LOL

# Q sounds like they're the same thing
# A they are
# Q so why do we need to convert an image into a latent vector?
# A we don't

[strange mist]

🙃

[hardy cipher]

thos conversation is really top notch to be quite honest

[strange mist]

is that from the comfy VN tutorial? 🙂

[high skiff]

RTX 4090 has higher FP16 T flops (83 > 61), faster VRAM (21gbps > 20gbps), more transistors (76.3B > 57.7B), more TMU's (512 > 384), more ray tracing units (128 > 96), more tensor cores, (512 > 192).

ROP's don't mean anything for perf, thats its capability of transferring what its texture mapping units (of which it has way less of) into a pixel output. More ROPs, the higher the max res the GPU can support, nothing to do with perf.

[hardy cipher]

oh, had a specific question for this AI, and it autosuggests both the questions and the answers, so just had it go off on a tangent. only change things up when it starts repeating itself

[indigo carbon]

so, are you telling me that if AMD would be as good for AI as NVIDIA, you would still give your money to NVIDIA?

[high skiff]

no, I would go AMD, but AMD is nowhere near NVIDIA right now for AI

the 7900XTX is considerably weaker in compute than the 4090

[strange mist]

oh, lol. crappy AI 😄

[high skiff]

even without software tricks

[indigo carbon]

I'm getting close to have an img+img AIT workflow =]

[hardy cipher]

oh, it's quite good at what it's supposed to be good at. but conversing me isn't one of it's strengths.

[high skiff]

@indigo carbonif AMD was even the same perf as NVIDIA in AI, I would have gone AMD, but its just not, and it never will be on this generation

[vital ermine]

Seems really weird and it persists across all. That only leaves some driver somewhere forcing it to run 100% or it is in the uefi of the mobo. No other thing could it be that is left.

[high skiff]

these are different drivers one ach GPU, spanning almost a year, on 2 entirely different PC's

I guess I can chaneg my drivers and see

[marble sapphire]

Is it okay to use clipdrop stable diffusion xl to generate image and use clipdrop clean up to remove the watermark and then use it for commercial?

[high skiff]

likely not. I recommend just using SDXL locally, or through this server in one of the bot channels., You can get much better results, and not have to fus with any of that

[marble sapphire]

Alright cool thanks

[high skiff]

alright, there is a slightly newer driver now, so I will try that

[indigo carbon]

also this still uses AIT while having the ability to fuse images together

[high skiff]

alright, drivers updated

@vital ermineDrivers didn't change anything

[vital ermine]

I don't mean gpu drivers

[high skiff]

this is so annoying, I need to know why this is happening. Its a serious problem for me

[vital ermine]

have you done a nvidia-smi while idle and it is max ghz?

[high skiff]

what is SMI? I have heard of it before

[vital ermine]

open a cmd and type nvidia-smi

[high skiff]

alright, and what am I doing here?

[vital ermine]

it reports to you and is very powerful . screen shot it and show me

[vital ermine]

well, shit

you have a ton of shit using that card

[high skiff]

its only AI that makes it behave this way

close AI, it behaves normally

[vital ermine]

only 1% for that snapshot in time

[indigo carbon]

CUDA11.8>>>>>CUDA12.2

[high skiff]

I have had this debate several times

it really does not matter

[vital ermine]

wdym close ai?

[heady vale]

what browser are you using. tried a different one?

[high skiff]

close CMD with any AI program loaded. Comfy, Auto, Kohya, Derrian, RVC, Ooba, Silly Tavern

only difference is I closed the comfy UI CMD

[vital ermine]

if you close all the ai out and it goes back down then that should tell ya what is doing it. How to help that I dunno

[high skiff]

its only for AI

[vital ermine]

right onboard internal gpu I know

[high skiff]

for whatever reason, if I have any AI program loaded in CMD, it forces my GPU to max clocks while doing nothing

[high skiff]

no, I am saying this issue only happens for AI. Nothing else on my whole PC causes this

[vital ermine]

then pytorch is doing it

[high skiff]

Firefox, and no

wait, I think I found a post about it

its for linux tho

for me, my idle goes from like 25 to 130

which is a little bit more of an issue than 7-20

ok, another detailing something similar on linux

[soft zealot]

silly question but have you done a complete clean OS install (possibly on a seperate drive) ? If so doies the issue persist?