✨｜sdxl

[uncut fiber]

you want modern photography of a woman or no colors?

[noble shoal]

The first Color Photo was taken 1861. 🤔 I mean I guess it follows the prompt, but was that really what the initial prompt from qwerty asked for?

[uncut fiber]

i think czech inventor made it

[ionic dragon]

it should depend on the prompt, i am collecting these prompts, so I can make the prompts I gathered from multiple hf datasets and then train on them

[uncut fiber]

o.k. it isnt low light and dark....

[ionic dragon]

whats the prompt?

[uncut steeple]

Took the last output and "Can you make it a short prompt, separated with commas"

[zinc cargo]

antique photograph of a beautiful woman, sad eyes, sad smile, facial symmetry, frontal view, cracked and faded photo paper, staring at the camera, headshot, dark background, 1850, low light, dark, 4k
his prompt with a lora that i had on

[zinc cargo]

all with protovision from @delicate kelp

[uncut fiber]

i have them in frame, why!

[ionic dragon]

left(original prompt), top-right(my modified prompt using llm), top-bottom(gpt prompt by boto)

[zinc cargo]

no lora

[floral island]

uuh... interesting wardrobe

[ionic dragon]

only this?
and yet you got great prompts
can you try this highly detailed exquisite warframe fanart, worms eye view, looking up at a giant 500 foot tall beautiful saryn prime female warframe, as a stunning anthropomorphic robot female dragon, sleek smooth white plated armor, unknowingly standing elegantly over your view, you looking up from the ground between the robotic legs, nothing but a speck to her, detailed legs towering over you, proportionally accurate, anatomically correct, sharp claws, two arms, two legs, robot dragon feet, camera close to the legs and feet, giantess shot, upward shot, ground view shot, leg and thigh shot, epic shot, high quality, captura, realistic, professional digital art, high end digital art, furry art, macro art, giantess art, anthro art, deviantart, artstation, furaffinity, 3d, 8k hd render, epic lighting

[uncut steeple]

Warframe

[ionic dragon]

you only got a single word?

[uncut steeple]

No lmao this was my comment

[noble shoal]

Shot shot shot shot Warframe. The LLM gets PTSD from that

[floral island]

haven't done llm prompts in a long time 😮

[uncut steeple]

Highly-detailed Warframe fanart, worm's eye view, 500-foot tall Saryn Prime, anthropomorphic robot dragon, sleek white plated armor, elegant stance over viewpoint, ground view between towering legs, proportionally accurate, anatomically correct, sharp claws, two arms, two legs, robotic dragon feet, camera focus on legs and feet, giantess shot, upward shot, ground view shot, leg and thigh focus, epic perspective, high-quality, Captura, realistic, professional digital art, high-end rendering, furry art, macro art, giantess art, anthro art, potential feature on DeviantArt, ArtStation, FurAffinity, 3D rendering, 8K HD render, epic lighting

Idk doesnt seem much different

[uncut fiber]

those old photos benefit from higher cfg

[floral island]

i mean, how hard can prmopting be? just prompt "fantasy art", no negatives. easy. KAPOW, now i'm an artist too 😄

[noble shoal]

You are a true prompt magician.

[floral island]

lets improve it 'generic dull fantasy art'

[uncut steeple]

LLM Prompt for fantasy art : Enchanted woodland setting, ancient towering trees, whimsical creatures, ethereal glow, mystical pools of water, reflections of hidden realms, vines intertwining, blossoming with magical flowers, soft beams of sunlight piercing through dense foliage, floating islands, ancient ruins reclaimed by nature, forgotten statues with eyes aglow, a central majestic ancient tree with secrets, tree homes nestled high, fluttering fairy lights, velvet moss blankets, curious mystical artifacts, shimmering leaves, a guardian spirit of the forest, subtle sparkle in the atmosphere, the dance of the spirits at dusk, surreal, whimsical, transcendent lights, soft palette with bursts of color, gentle magical aura, hidden pathways leading to mysterious adventures, storytelling through scenery, details that entice lore exploration, high fantasy, endless imagination captured in a frame, a scene whispering ancient lore, dreamy haze, a glimpse into a forgotten magical world

[floral island]

running a local llm?

[uncut steeple]

Nah chat gpt 4

[floral island]

generic dull fantasy art

[uncut fiber]

The A image is SD classic layout

[ionic dragon]

left(the original prompt), right(gpt modified prompt)

[uncut steeple]

I mean, not terrible

[uncut fiber]

iam artist too! Shorter sword would probably work better.

[noble shoal]

@ionic dragon Yesterday you shared a hugging face page with 1.82 Million lines of prompts. Don't they work?

I mean for training.

[noble shoal]

That is one sus Tail between the legs.

[floral island]

all the way from the top tho

cyberdwarves!

[hardy cipher]

whoa, what's he getting ready for in the bottom right image?

[crisp owl]

They call it "war"

[hardy cipher]

well he's definitely in good shape. can't deny that

surely excels in war activities

[crisp owl]

It's quite the exercise

[ionic dragon]

yeah pretty impressive, if we can tweak the prompt, then I think we can make it great

[hardy cipher]

decorated soldier that made it through the infamous bath house battles

look at this stupid cfg. not a world record but it's high

[hoary saddle]

anyone see any issues with installing cudnn_ubuntu2204-12-2-local? i recall some issues with certain versions of cudnn

[hardy cipher]

I don't see any issues and also don't know anything

[ionic dragon]

@uncut steeple can you try this
"generate a modified stable diffusion prompt of this"

[uncut fiber]

i was surprised how old qr monster making qr codes unreadable... But it is not main point of extension probably.

[noble shoal]

That's what Bing has to say to this:

You are standing on the ground, looking up at a giant 500-foot-tall Saryn Prime female Warframe. She is an anthropomorphic robot female dragon with sleek, smooth, white-plated armor. You are nothing but a speck to her as she unknowingly stands elegantly over your view. Her detailed legs tower over you, proportionally accurate and anatomically correct, with sharp claws. She has two arms, two legs, and robot dragon feet. The camera is close to the legs and feet, capturing a giantess shot, upward shot, ground view shot, leg and thigh shot, and epic shot. The artwork is of high quality and realistic with professional digital art, high-end digital art, furry art, macro art, giantess art, anthro art. The artwork can be found on DeviantArt , ArtStation , or FurAffinity . It is a 3D 8K HD render with epic lighting.

[ionic dragon]

oh cool, lemme use bing too

[uncut fiber]

It is a 3D 8K HD render with epic lighting. 😄 And best picture ever been created 😄

[noble shoal]

🤣 I didn't even read through

[uncut fiber]

best picture ever created i just added 🙂

[noble shoal]

Would be fitting

Well, can someone try this prompt? Now I am curious

[uncut fiber]

copax timeless

[noble shoal]

Ok, I googled Saryn Prime Warframe and have to say: Pretty inaccurate.

[indigo carbon]

am I using this right?

[ionic dragon]

does it work?

[indigo carbon]

it seems to.. I implemented AIT to it

[long oriole]

can you tell me what the name of this plz ?

[uncut fiber]

check readability @indigo carbon where i tried it percentage was really low.

[uncut fiber]

i meant readability of QR code generated

try those 🙂 @indigo carbon

[lusty wolf]

FreeU and Dynamic Thersholding.

[woeful patio]

anyone here have experionce with mangio-rvc I can ask a quick question?

[spring fulcrum]

I am sad today... I just found out that one of my favorite developers has passed away. They went by the name capybara_paws... They will be missed. 😦

[hardy cipher]

I keep reading about it but haven't kept up. have you tested them? I've read about how they work in the past. middle of the squares are really the only thing that matters ultimately, at least for the data portion. then there's other portions that need more precision. and then you can adjust error allowance as well. but it's pretty impressive how much can be taken away and still have at least some devices pick it up and read it

[uncut fiber]

i tried them on some qr codes checker. And what i supposed was easily readable was not recognized. Only 2 absolutely similar to original was read o.k.

[steady grove]

qr codes through controlnet. i think you should use a1. it's more nimble than comfyui. you won't have to rewire your whole ui and wonder if it's right, each time you try a different approach

[uncut steeple]

not bad tbh

[uncut fiber]

i am using A1111 with qrmonster, but there is different for qrcodes, i think more serious than creative

https://github.com/missionfloyd/webui-qrcode-generator

[indigo carbon]

the preprocessor turns that into a black image. what did you do to turn it into lines?

[vale eagle]

freeU + v hand lora (not released). It kind of interesting in last one. The ori does right but freeU does wrong.

[uncut fiber]

i no preprocessor. sd thing without preprocessor. Or you can probably use preprocessor invert colors?

more fingers more victory

[uncut fiber]

@indigo carbon it is black with alpha which is represented basicaly as black image.

[indigo carbon]

idk, I feel like I'll back off from image blending and zero subject generation until an SDXL version of BLiP diffusion releases. BLiP diffusion seems like the only method of controlling SD without loosing its talents

[indigo carbon]

I feel like FreeU actually makes results worse sometimes.. were you able to use AIT with FreeU or do we need new modules for that?

[vale eagle]

I noticed freeU sometimes make things worse but in general it give me good shape and details with a lot of extra noise. I am not using AIT for a long time because I am facing a lot of issue when I keeping everything is up to date.

[indigo carbon]

yeah, I get that. I kinda feel like we need an equivalent of exLLaMa for diffusion to make things easier

I tested exLLaMa recently and it was over x9 times faster than pure pytorch while even increasing quality slightly

shit went from 8T/s to 62T/s

[hardy cipher]

I honestly haven't had great results with Freeu, but might just be the images I've tried to make. not saying it doesn't have a positive impact on other stuff

[indigo carbon]

we can all agree that we need an exDiffusion thingy though, right?

[hardy cipher]

well if it's 9 times faster that sounds reasonably decent to me

[shy kelp]

is there a scribble cn for sdxl atm?

[vale eagle]

it good at shape and structure but bad at color and smooth some details. I used it with 1,1,0.9,0.2 with Crystalclear 3m sde karras

[urban totem]

What is the general feeling with the "SDXL 1.0" Vs "SDXL 1.0 0.9vae" debate, I haven't played much with SDXL and I'm wondering which one I should use

[indigo carbon]

yeah, autoGPTQ did about 8T/s on a 13b model while exLLaMa did over 62T/s on the same model. that's a much bigger boost than what AIT does. It would sure be nice to have something like that for diffusion

[noble shoal]

There's a debate?

[hardy cipher]

I tried it on a "vintage photograph" prompt and it bumped the contrast up to a pretty unpleasant level. but I realize it's not going to be one size fits all

[urban totem]

I meant the general take on it, I see SeargeXL has the 0.9vae one in its workflows by default and was wondering if I should download this one or the "normal" one

[noble shoal]

The normal vae had artefacts in the beginning

[urban totem]

I saw that on huggingface but they seem to have fixed it since according to commit logs

[vale eagle]

Yeah, it has contrast issue that why I used it to generate the depth map and based on it to generate new image without freeU

[noble shoal]

Then they should be equal now. But just in case, try both with the same settings and zoom in. If you see weird red and green artefacts then use the other one.

[hardy cipher]

she hears extra good

[urban totem]

Alright, thank you!

[floral island]

remember to feed your waifus in time

[hardy cipher]

are you guys scared?

[noble shoal]

Nope

[hardy cipher]

why not?

[noble shoal]

Too cute. No scary concept. Sorry

[hardy cipher]

yeah. I was going to just post it then I considered the idea that someone might be offended by it

[spring fulcrum]

maybe these are a bit more scary... the nun 2 is out

[noble shoal]

In this case I am going to report you 🤣

I think things without clear faces like shadowy dark slim figures in a dark and misty forest at night are a bit more scary.

[zinc cargo]

you got qr to work on ait?

[crisp owl]

looked like canny controlnet

[indigo carbon]

.

[noble shoal]

Now we're talking. Super scary.

[noble shoal]

For a moment I thought that you got SD to generate raindrops in front of a subject. Now, that would have been scary. But no, just behind.

[zinc cargo]

welll it's kind of beyond them...

[shy kelp]

i threw my pixar prompts into 1.5 and god it was awful

[shy kelp]

wildcarding is the same thing as if it was {like this|or that|or this|or that} right?

but it just pulls from a file or something?

[crisp owl]

in comfy?

[shy kelp]

ye

[crisp owl]

There is no pulling from files natively in Comfy, but the wildcard prompting otherwise works the same

{thing1|thing2|thing3|thing4}

[shy kelp]

i queued up like 20 images earlier today and just had brackets with different things and it just output a bunch of random images (all expected tbh) and was wondering if there was a simpler way to do it instead of the bracket way so was looking at wildcards

[crisp owl]

There is file calling from custom nodes, just not natively

[shy kelp]

so i could do like hair and it just pulls hair color options in hair.txt or whatever but was expecting something like

blue
black
green
yellow
auburn
etc..

[noble shoal]

Yes. There are nodes that do that. I have even written one for that myself.

[zinc cargo]

@JPS wa wa wee wa

[ionic gulch]

(all red/black/white horror girls are from a run of 100 pure text prompts)

[indigo carbon]

which CNET did you use here? canny? or something else?

[hardy cipher]

very classy

[hoary saddle]

comfyui rules, api is a way of life 🙂

vary 3 lvls of denoise

[noble shoal]

Got it blending

[uncut gull]

It's not just the card art, you know? It's every aspect of a game like this. Gatcha games are addictive because of their shininess. What if open-source, non-monetized games could compete?

[slender coral]

Could someone point me in the direction for a depth map node in comfyui, all I can find is for auto1111

[hardy cipher]

do you have the manager?

[uncut gull]

In A1111, go to extensions and search for "depth". Tons of features, including 3d mesh generation, stereo images, video, download models right from the tab, etc.

While gambling games are evil, they are also the highest-grossing apps in existence. If an open-source alternative existed, without monetization, maybe it could help people?

[slender coral]

Sorry in comfy (I don't know why I didn't add it there)

[hardy cipher]

there actually aren't any on comfy. sorry

[uncut gull]

But aren't there controlnet preprocessors? (Maybe there aren't, I really don't know.)

[hardy cipher]

nah, but this guy doesn't like answers and seems to normally ignore people that try to give them

was going to tell him how to find them, but then he ignored me to repeat his issue to you

very strange psychological stuff going on there

[slender coral]

Arn't you the lad that said 0.5 denoise changes a lot?

[glad grove]

he prob thinks we are some sort of a personal costumer support here

[slender coral]

I don't really trust much of what you say 😄

[uncut gull]

Count me out of drama, thanks. I'll stick to prompting SDXL.

[hardy cipher]

why would you have a screenshot of something I said days ago?

[slender coral]

I just took it, why did you give wrong advice? Then complain when people don't "listen"?

[hardy cipher]

I think I'll put him on ignore. nothing gained

[glad grove]

i too have several screenshots of my sworn enemies on a folder called "potential ammo

[hardy cipher]

I mean, I was trying to help the guy. he had a very low contrast image he was trying to do something with, and those can change quite a bit with a little push

[slender coral]

No ones forcing you to answer anything. And I'm not expecting anyone to answer, that's the beauty of discord.

[hardy cipher]

just some weird stuff going on

[slender coral]

I appreciate you trying to help but you don't give good advice, it tends to be wrong, and now you're complaining that I don't respond. So sure, weird stuff. But thanks, feel free to block me.

[glad grove]

now speaking of clowns

[hardy cipher]

well anyway, I apologize for opening the can of toxicity here

[slender coral]

I switched to comfy after auto1111, was using that plugin in auto, how's the gen times?

I may just gen in comfy then jump into auto to get the depths.

[hardy cipher]

lol

[uncut gull]

I don't use Comfy, so I can't compare them. On my PC, Res101 gives the fastest depth maps, and the most accurate for the scenes I've tried so far. There are several models available, and they are all different sizes. You can also control the dimension it runs at for speed up.

[slender coral]

I'll jump in and test, thanks mate.

[uncut gull]

Yep. Those are menu buttons.

Okay. I'm all out of awakeness. Good night all.

[hardy cipher]

look at the enthusiasm in his face

[upbeat summit]

this image pops - cool contrast effect

[hardy cipher]

thanks, forget what I even did. just looking back through some things

[slender coral]

I've spend hours trying out different settings and think I have it pretty close to what I'm after, I just want to check here to see if there's something that I'm overlooking, do you guys see anything that would increase the detail in my high resolution that I may not be catching?

Different models, samplers, any denoise settings?

[hardy cipher]

I'm quite impressed by these strawberries so many quality strawberries

[uncut fiber]

it is sd qrmonster, sorry posting it here

[upbeat summit]

Nice composition and mood! Good high_res_upscaler choice ;). I can also recommend the s64w8 variant.

Detailed wide-shots are not easy. I worked on some aerial shots prompts and the resolution is just not there yet. But the detail in your image is pretty great.

Your pipeline has a couple of steps so I can't really say at first glance what I might tweak without working with it myself.

Depending on the style you're aiming for you could probably refine the sharpness. You could counter some of the slight blurriness (haloing) by increasing your denoising strength for your high-res fix / upscale pass - but this might also change your composition too much.

I don't know if you use a prompt but you could also try adding some negative prompt tokens to enhance sharpness and fidelity.

[slender coral]

the variant? https://github.com/JingyunLiang/SwinIR

[uncut fiber]

my favorite

[slender coral]

👍🏿

[upbeat summit]

https://openmodeldb.info/models/4x-classicalSR-DIV2K-s64w8-SwinIR-M

[slender coral]

Quick explain, what would be the difference between medium and large?

Is large better?

[uncut fiber]

could be in for what Xx it is trained?

[slender coral]

1.0

[uncut fiber]

models on openmodeldb.info are trained on different upscale, which shouldnt be changed. 1x 2x 4x 8x 16x
can imagine this is the main difference

1x upscale, no upscale is for restoration purposes

[upbeat summit]

Afaik SwinIR-Large is a newer version and trained on a larger network. That doesn't imply that it's better but different - depends on the use-case. I'm not a huge fan of pixel upscaling in general but wanted to do some experiments.

4x-Ultrasharp is used quite a bit and the performance is really good. But after testing 20 - 25 upscalers I found that SwinIR s64w8 gave me better results - less over-processed artifacting and cleaner sharpness while only being slightly slower.

[hardy cipher]

not what I expected

[glad grove]

he looks so healthy 🤣

[hardy cipher]

he has a large brain

[glad grove]

new pet

[glad grove]

now thats a cute couple

💏

[hardy cipher]

think that might be ET under there

[glad grove]

he needs love too

[nimble heart]

makeup a little whiter he almost lookes like Sweet Tooth

[nimble heart]

why is the ground made of the fast travel rails from Sonic the Hedgehog

[hardy cipher]

just how they live. they move fast

[nimble heart]

the next mad max movie should 100% have an alien fish person invasion that unites all the clans

[hardy cipher]

activating maximum cfgs

[urban fjord]

I'm testing out a model I'm training and these kind of looks like convincing movie posters

(Ignoring the bad text)

[hardy cipher]

indeed. it's that one lady

[uncut fiber]

dear XL friends can you stops in General with images and told me which hidden lady has higher subjective quality?

[uncut fiber]

It reminds me something i already saw

[hardy cipher]

did you see that? or something similar?

[uncut fiber]

yes, this not. I recall eyes with skulls from somewhere

maybe similar

[vital ermine]

I find it weird that more steps in a gen changes the images almost completely

[urban fjord]

Knock-off Star-Wars

[urban fjord]

That's how Eular a and other "a" samplers work

[nimble heart]

SDE can change it a little too but not as much

[hardy cipher]

I think it depends on the model

ahh, I should have scrolled

[nimble heart]

not really

[hardy cipher]

well some just make it more of whatever, some alter it completely. the ancestral ones

[uncut fiber]

do not too much samples in some samplers spoile image?

[nimble heart]

also all samplers will rapidly change before they converge if your step count is too low

but after like 30 or more steps the changes should be small all the way up to 100+ when you use a deterministic sampler

[vital ermine]

this is not an A sampler

[hardy cipher]

heun don't change. heun heuns

[nimble heart]

heun is lame

[glad grove]

heun my beloved 💖

[vital ermine]

DPM++ 3M SDE Karras changes, but all of the DPM do

[nimble heart]

SDE

SDE introduces extra variance in each sample

[hardy cipher]

what I like is that it takes twice as long per step to make things without as much variation

[glad grove]

Heun ❤️‍🔥

[upbeat summit]

Deterministic, non-ancestral samplers are stable. ex. try dpmpp_2m with a normal scheduler

[nimble heart]

apparently adding the extra stochastic noise or something means it can self-correct I guess?

but yea the tradeoff is you dont fully converge, the image will drift away when you add enough samples

[vital ermine]

all I have is the sampler

[nimble heart]

why though

what happened to your other samplers?

are they okay?

[vital ermine]

not in response to you was to mass

[nimble heart]

scheduler shouldnt affect determinsim I dont think

it might converge differently but it'll still converge

[upbeat summit]

yeah - it just deviates a bit more

[upbeat summit]

was just an example. you can use other samplers and schedulers of coursse

[uncut fiber]

i like Ancestrals because sooner cleaner flat background.
Probably because i have not powerful gpu enough

[nimble heart]

karras seems to generally converge slightly faster

but sometimes I found it can kinda get sidetracked almost? So @ low steps like say 20 it'll be more different than a normal scheduler

[uncut fiber]

i think it was at least 30 steps 32 i usualy use. Not sure, datas are in picture

[upbeat summit]

I'm currently using dpmpp_3m_sde_gpu karras between 55 - 75 steps, CFG 3.5 - 5

[nimble heart]

SDE's rate of change seems to slow down, so I guess it kind of converges once you hit like a few hundred steps

normal 20, normal 40, normal 100
sde 40, sde 100, sde 300

[upbeat summit]

yeah running SDE for too many steps can also have some side effects - making the image overly smooth

[uncut fiber]

How about restart, should be good

[nimble heart]

good ole 2M karras is probably my favorite next to DDIM

[uncut fiber]

@nimble heart have you tried Restart? It is available for comfUI as well

[nimble heart]

"restart"?

what's that

[uncut fiber]

It is SAMPLER ! 🙂

this is in it. It is somehow able to get back probably therefore Restart

It is A1111 but tried it in Comfui as well, i think some special Ksampler for it, not sure now

[nimble heart]

me no see

[uncut fiber]

check for ksamplers, or in manager check for restart

i am in A1111 and dont want to leave it.

[nimble heart]

I dont have the manager

hey this one's new

ddpm

[uncut fiber]

https://github.com/ssitu/ComfyUI_restart_sampling

[upbeat summit]

has been for some weeks

[nimble heart]

damn unipc converges kinda fast

least on the subject

[upbeat summit]

unipc is interesting. behaves very different

[uncut fiber]

@upbeat summit did you try restart?

[misty sundial]

Hi I'm trying to use aspect ratio helper to resize my picture from 1024x1024 to 1920x1080, but the resized image comes out distorted. Is there any methods to make it so it doesnt get stretched?

[nimble heart]

yea I never used it much cause it used to not work on AMD

[prime bison]

Yeah, use supported resolutions

[upbeat summit]

I tried it a couple of weeks ago, but didn't do enough testing. since it doesn't support SDE and that's what I'm currently using, I haven' t looked into it again

[uncut fiber]

1344x768 probably best for you

[nimble heart]

back when I had to build PyTorch myself I didnt know how to include hipmagma over whatever the default cuda implementation is

[upbeat summit]

ah didn't know that

[nimble heart]

it works on the official binaries

[prime bison]

Get as close to 1040000px as possible in general for the generation. @uncut fiber suggestion 1344x768 is probably best

[uncut fiber]

it comes back and last quite longer

[nimble heart]

but if you build it yourself it needs magma which idk how to bundle

ddpm is a goofy mf-er

[uncut fiber]

resolution SDXL was trained on @misty sundial

[misty sundial]

I'm fairly new, so I'm not sure how to do this? do you mean to upscale the picture first to like 4k x 4k?

[prime bison]

No, I mean don't start out with 1920x1024. SDXL base requires ~1mp total, so use one of the trained resolutions

[nimble heart]

aight this shit's too complicated imma stick to the builtins

[prime bison]

You can upscale after that if you want 1920x1024

[uncut fiber]

render at resolution it was trained. all is around 1Mpix And when rendered resize it
yes

[nimble heart]

when they start adding the UTF-8 chars my koala brain cant keep up

[prime bison]

Isn't that just italics? Looks like italics to me.

[misty sundial]

So I should be creating the initial image in my desired aspect ratio of 16:9 and then resize it?

[prime bison]

Real question is why are they using emdash for subtraction

[prime bison]

Yes. For 16:9 idk the correct resolution, it's whatever the bots are using

1344x768

Which was already suggested

[misty sundial]

Yes I see that from the file that @uncut fiber posted! ty

[nimble heart]

1368x768 is closer to 16:9 and works well too

[misty sundial]

But if I already have an image of say size 1024x1024 is it not possible to resize the ratio?

[uncut fiber]

it is not my work all credits go to creators 🙂

[prime bison]

Basically if the bots support the resolution, it should work. Though I've had issues with some of them on SDXL beta that's a different issue.

[crisp owl]

with something like outpainting, in a sense you can "resize", but eh

[uncut fiber]

it is possible. Strech it, or add black bars. To not it be distorted

[misty sundial]

do you mean to use outpainting to recreate additional background to widen it? Is there a guide you have in mind, I think I might have watched something like that before

[crisp owl]

I've only used outpainting a single time just for the curiosity sake back in 1.5 days, so I'm unsure what SDXL resources are available for it

[prime bison]

I never managed outpainting. Never got it to actually work on Colab, and Horde supports it but none of the UIs I can use do (Android or web)

[uncut fiber]

openoutpaint is nice. For A1111.

[prime bison]

I tried like 3 different ways with a1111, including proxying whatever that website was inside colab to get rid of cross origin issues, didn't work 😦

[uncut fiber]

this is all in openoutpaint. It is very creative. Starting from 0. But yes, lot of color inconsistency, but it is fun create that way.

[zinc cargo]

is there a discord server of channel dedicated to ComfyUI?

[prime bison]

I can see your borders

Lol

There's multiple discords dedicated to ComfyUI

[uncut fiber]

@prime bison whose borders?

[prime bison]

Search Reddit, I found like 4 ComfyUI discords last time I did

[prime bison]

I see what looks like a square off to the left

[uncut fiber]

it has not borders, yes probably some blend issues, but it is generated from 0 i started in center with say 240x240 and build up on it with so smalle squares. So it is not extended image by outpainting it is created from 0 🙂

[prime bison]

You can tell when outpainting was used sometimes if they didn't do a wide enough blend

Yeah

It's the squares then

Oh, you did the whole thing in outpainting?

[uncut fiber]

yes

[prime bison]

Ugh. I g2g. Washer demands my attention or it overflows

[crisp owl]

gl

[uncut fiber]

gl

[nimble heart]

interestingly, it only looks like it converges faster on the tiny preview. The details converge less. maybe it has noise like SDE?

[prime bison]

UniPC has some kind of correction built-in IIRC

Stupid clogged pipe

Unified predictor corrector

[fierce turret]

does automatic1111 work with sdxl

[crisp owl]

yeah

[prime bison]

UniPC sampler is a method that can speed up this process by using a predictor-corrector framework. It predicts the next noise level and corrects it with the model output

[fierce turret]

i have latest dev branch and it takes more than 20gb dram (ddr5) to load the model

idk why

[prime bison]

Use ComfyUI instead

[crisp owl]

I stopped using A1111 and use Comfy

but, it does work-ish

[prime bison]

I don't use local at all, and I've figured out that much lol

[fierce turret]

i mean fair

the sphagetti nodes

[prime bison]

There's other UIs built on ComfyUI

[fierce turret]

oh lemme c

[prime bison]

Idk the names, but they do exist

[crisp owl]

#🐝｜swarm-ui

[fierce turret]

oh ok thx

[prime bison]

Isn't stable swarm for queueing or something?

[crisp owl]

It's a UI with Comfy backend

McMonkeys project

[prime bison]

True, but IIRC it's not general purpose. Could be thinking of something else though, I don't actually follow this stuff, I just stick with the bots here and Horde for everything.

[nimble heart]

stable swarm is a pretty simple ui in comparison to a1111. lacks extensions and such

[prime bison]

Clearly thinking of something else then

[crisp owl]

Alex (mcmonkey) — 08/17/2023 2:11 PM
you should have access to everything, barring some extensions have compatibility issues (that are easily fixed just need to identify which ones + report to them)

[urban fjord]

ComfyUI is only worth it if you're using the AITemplate stuff to speed things up

[nimble heart]

comfyui is also like 30% faster on my card than equivalent settings in auto

[hardy cipher]

or if you like putting things in different configurations than a1111 provides

it would be so tedious to do a lot of the things I do on comfy with a1111

[nimble heart]

in ComfyUI my max SDXL speed for 1024 is 3.17 it/s
auto was like 2.52 it/s

[hardy cipher]

might be possible, but not really worthwhile

[prime bison]

The little that I know is that ComfyUI also enables far more complex workflows than a1111 or other UIs do easily

[urban fjord]

Auto1111 enables workflows that I actually care about while ComfyUI doesn't support the workflow I care about and enables a bunch I don't care about.

[hardy cipher]

let me tell you, it's really not that hard either. you have to connect a few things that have the same names and are color coordinated

model output to model inpjut

[prime bison]

I'm unaware of anything being possible in a1111 and not in ComfyUI....

[hardy cipher]

clip output to clip input

[prime bison]

Other than the use of specific a1111 extensions (which obviously require a1111 to use)

[nimble heart]

iterative inpainting is more tedious in comfyui and afaik there's no easy openpose editor node but thats about it as far as major complaints go.

[hardy cipher]

a1111 certainly has a few extensions that aren't on comfy, so there is that

but I also don't get the UI wars some people seem to feel they're in. it doesn't have to be a binary one or the other thing

[prime bison]

I won't touch ComfyUI anyways, far too much for the little bit of stuff I do, and I don't even have a GPU anyways. 🤷

[hardy cipher]

it's really not though

[urban fjord]

SwarmUI added proper x/y/z support so that's less of a sellingpoint for Auto1111 now, but ComfyUI without swarmUI is quite limiting.

[nimble heart]

ah grids too

thats the other thing comfy misses

[hardy cipher]

there are grids

if you want ot use them

[prime bison]

Having movable nodes is more than I need

[nimble heart]

wer

[crisp owl]

Efficiency Nodes has an XY script

[hardy cipher]

https://github.com/LEv145/images-grid-comfy-plugin

[prime bison]

I do oneshot with Lora and TI and hires-fix with whatever single model and I'm happy. That's it.

[hardy cipher]

well that's fine if it works for you

[urban fjord]

That extension is making me want to vomit.

[nimble heart]

yea nah

[hardy cipher]

I don't use it, lol

[nimble heart]

way too much work

[hardy cipher]

I just said it's there

[nimble heart]

its easier to just queue up everything at once then combine the images yourself using magick's grid feature

[prime bison]

montage -mode Concatenate -tile 3x3 ?.png out.png my XY charts

[nimble heart]

exactly

I dont wanna rebuild my whole node layout to make a grid testing a setting

[urban fjord]

I don't know if SwarmUI supports different LoRA and ControlNet in its X/Y/Z grid stuff.

[hardy cipher]

well whatever, it wouldn't be that hard to just make grid nodes

[glad grove]

since comfy doesnt have a preview for loras i have in folder i cant use it ( i have 2300 loras ) but for speed alone it smokes auto1111 and its memory hogness

[hardy cipher]

don't think I will

[nimble heart]

grid shouldnt be a node

[hardy cipher]

well what should it be?

[prime bison]

If I don't use only ~9 images (I did 121 earlier) I just need to make sure the generation id counter is at the front of the filename so it orders properly

[west breach]

you'd have to run all the images in the one workflow to make a grid, otherwise how could you do it?

[nimble heart]

grid should be a menu where you can target specific values for the XYZ
So on the default workflow you'd right click the CFG value on ksampler and "set grid X" or something, then in the grid menu you can put in values

like how blender you can set node values to be drivers or something

or animate them

[hardy cipher]

I would like that, but I don't know if it'll happen any time soon

[nimble heart]

so any widget based node field would be grid-able

[hardy cipher]

I was just thinking in the context of custom nodes

[nimble heart]

nah

adding a bunch of nodes just to compare cfg is lame

[prime bison]

I use horde/ArtBot. Downloads come with metadata. The default to order by finishing time, not by starting time, but I've got a script to rename by start time and another to rename by other contents so they can just be sorted automatically if I didn't generate in order or one never came back.

[hardy cipher]

well I don't know how else I'd make an alteration to the program

[nimble heart]

again at that point its easier to just magick montage

[misty sundial]

Does openoutpaint still work? I'm having trouble trying to install it, I get the error PermissionError: [WinError 5]

[hardy cipher]

not saying nodes would be the most efficient approach

[uncut fiber]

you have to add in commandline args in webui-user.bat --api

oh install it... i thought run it

[vital ermine]

has anyone made a scripts node for comfy? As a trainer that is invaluable so I am forced to use Auto due that.

[urban fjord]

What do you mean by a scripts node?

[vital ermine]

X/Y/Z Plot specifically

[vital ermine]

Kick ass

S/R

[uncut fiber]

fight fight

[vital ermine]

Now to switch over to that and give it a whirl. Been fighting Kohya all damn day as it is refusing to act right with the training.

[urban fjord]

What's the problem you're having with Kohya?

[crisp owl]

I tried earlier today to do a very basic training in Kohya for Vivi Ortinier from FFIX, but my training didn't do anything lol
But I ain't doing anything more, my 2060 can't handle it

[vital ermine]

I switched to a different trainer and wanted to see how Kohya would react if better and Kohya just refuses to train it. OneTrainer had it trained in under 1 epoch. Being epoch based I wanted to try Kohya for step base as 1 epoch was slightly over cooked

[urban fjord]

Yeah, 6GB VRAM is too little.

Better just train in the cloud at that point

[crisp owl]

It's the Super model, so 8gb, but still yeah, it's still a pretty small amount nowadays

[urban fjord]

Oh, you're running with a ton of repeats.

[vital ermine]

My 1060 couldn't train anything except embeddings and it took 30s per step

30-50s

100 steps was 10-20m

repeats are evil

OT doesn't use them

[urban fjord]

It's just strange that you overcooked it at 1 epoch.

But I guess it depends on dataset and learningrate.

[vital ermine]

you don't know OT very well. That thing rocks

[urban fjord]

And if it cooks at 1 epoch your learningrate is probably too high.

[vital ermine]

has some thing called rex and it is so fast in how fast it trains it is crazy. I couldn't even slow it down and needed to

[urban fjord]

OT can't do magic, it still has to follow the same rules as everything else.

[vital ermine]

tried the LR but no dice

Oh, well, I am super happy to give the middle finger to Kohya scripts finally so was just seeing anyway.

[urban fjord]

Controlnet requires lower LR and higher epoch it seems like so now I'm regularly up at 1000+ epochs with Kohya-ss.

[vital ermine]

the LR for Rex is at .00038 I think

[urban fjord]

While for LoRA I tend to stick at 80 epochs.

Yeah, that is about the learning rate I'm using with Controlnet atm.

[vital ermine]

I tried 1e-4 and still not slowed down

I am still not sure how rex works but they are adding adafactor after I showed them how good it is. They really liked it.

that will help

did you know we are about to get Xformers V2 which conserves more memory and is faster too? I can't wait and I hope it comes with Pytorch 2.1

poor ole SDP was a dud

[glad grove]

lemme guess only works with RTX

[uncut fiber]

@upbeat summit with lora?

[vital ermine]

works with whatever Xformers works with

[uncut fiber]

hope it will works with AIT as well

[vital ermine]

https://github.com/facebookresearch/xformers/issues/795

Don't even care as AIT is dead to me

[upbeat summit]

yeah - for the glitch style. working on demo images for Aether Glitch by @icy brook right now

[glad grove]

FlashAttention-2 currently supports: Ampere, Ada, or Hopper GPUs

[vital ermine]

LOL

well, on the github they said Pascal too

[uncut fiber]

nice 🙂 @upbeat summit

[vital ermine]

Guess turing but rtx is turing so wtf?

[glad grove]

yea hopefully a kind soul shows some love to pascal 🙏

[urban fjord]

If it weren't for the spelling/text, I could almost be convinced these were real posters...

[vital ermine]

yeah, but see if it is an rtx issue then Turing should be listed as well so it is something else I guess

[hardy cipher]

come at me, bro

[uncut fiber]

spider cat 🙂

[hardy cipher]

they're special

[vale eagle]

I have an idea about make use of sd1.5 for sdxl. Sd1.5 has various community resource but sdxl couldn't use. How about using sd1.5 to generate depth map and use sdxl based on the 1.5 depth to generate the image. It could make use of 1.5 resource to get better hands and totally generated by sdxl for realistic and color tone.

[hardy cipher]

well I don't think the preprocessors are dependent on the version are they?

[crisp owl]

nope

[urban fjord]

The preprocessors are completely indepdenent from the SD version.

And SDXL seems to have better hands anyway so not sure why you'd use 1.5 for that.

[hardy cipher]

I made one. it might not be good, but it's mine

[vale eagle]

hands just for an example. The point is make use of 1.5 resource

In some case, sd1.5 still win

[glad grove]

the better hands

[urban fjord]

Controlnet won't help if SDXL is completely clueless in how to make it.

[crisp owl]

when did he join ICP?

[supple knot]

Is FreeU any good?

[crisp owl]

I haven't seen anything extremely compelling from what examples I've seen people showcase

[supple knot]

its suppose to give finer detail

[hardy cipher]

it can

[crisp owl]

It alters the initial training weights of SDXL. There's nothing necessarily in that that "gives finer detail".

There's claims on what it does, but I see it being treated as basically more variables to toss into the woodwork

[hardy cipher]

I'm sure people will find interesting uses for it. I was in here when they were talking about it right before it was released

[misty sundial]

I was running a generation, but then I interrupted it. But I think artifacts are left on the gpu, because I w tried for another run it says im out of memory. Is there a way to clear the gpu again without having to restart SDXL?

[hardy cipher]

nothing wrong with it. just don't think anyone has found the ideal usage for it yet

[vale eagle]

In my testing, freeU could generate "better" structure but it has contrast issue. It is not generally good but it could be used in some case

[crisp owl]

freeU conversation started here
#✨｜sdxl message

[crisp owl]

not that I've found.
I run mine in powershell so I can just ctrl+c my instance, then just hit up enter to restart.

[indigo carbon]

Honestly, I feel that. I think the ultimate solution for optimization would be an equivalent of exLLaMa for diffusion

exLLaMa can run LLaMa about 8 times as fast as pure pytorch, which is a much bigger boost than AIT does

Also exLLaMa is more flexible as it's not architecture specific as far as I know

So if we could get an exDiffusion that would be great

[vital ermine]

Agree

AIT is very inflexible and one comfy update it blows it all up

besides BS4 max is another issue with it.

plus I couldn't get refiner to work with it it had a recursion error no matter what I did

[noble shoal]

I still have an annoying waiting time before inference when I use Loras. Makes the faster speed pointless. But yeah, 12 gigs of vram seems to be the bottleneck.

Got it mostly out of my workflows.

[fierce hollow]

controlnets are completely broken with ait for me

even if I'm not using ait it's patching the loader and breaking something

have to manually rename the whole folder to .disabled for it to work

[ionic dragon]

@crisp owl can I dm you?

[noble shoal]

I think the "T" in AIT stands for "Trouble"

[uncut fiber]

@noble shoal just disable keep in something. It is only trouble i always forgot switch it and must restart

[vital ermine]

I have 48gigs and iffy

LOL at the T

[noble shoal]

You Monster 😱. Is that a quadro or something?

[vital ermine]

oh, you said ram not vram. I wish I had 48gig of vram

[noble shoal]

Pfhuu. For a moment I was very jealous

[vital ermine]

Don't be because to really train XL you need 80GB of vram. Not joking as it takes 73ish gb. Forgot what he was saying but if you lower the BS down for FT with a large dataset 43gb.

slow as in days for them he said

rent an a100 and it was sweet

80gb

[glad grove]

more than enough for a H100 then

[noble shoal]

Oh them sweet a100's. I wish Nvidia would make something like that for normal people.

With a regular income

I mean, I don't need it for gaming

[vital ermine]

"I've been training a generalist (but mostly photorealistic) model for... 3 days now on a A100"

Shit me for 3 days and counting no way

try the model and fail and just oof

I have a lora being a little bitch all day

something is not training then when it does it is iffy

[glad grove]

its only 27k euros https://smicro.eu/nvidia-h100-94gb-hbm2-900-21010-0020-000-1

[noble shoal]

Thanks. Going to call the bank.

[vital ermine]

anyone here train loras?

[noble shoal]

Kinda

[vital ermine]

Is it the TE that prevents my lora from working if I say X in background but leave that out it works?

[noble shoal]

I lost you at "TE" 🤣

[vital ermine]

I despise SAI for this dual TE nonsense as no trainer is overjoyed with their nonsensical move. Plenty of work to eliminate it coming.

[noble shoal]

Wait, now I got it

[vital ermine]

G and L

[floral island]

wait, are we flexing available gpu power? 😮

[vital ermine]

not anyone in here is with XL

[glad grove]

no,im stealing it from some server

[floral island]

I'm stealing gpu power from my boss (When our servers are not in use)

[noble shoal]

Yes, general awareness has a Homeserver with 1 TB vram at home

[vital ermine]

LOL

damn I wish

[glad grove]

he has money but i have a gun ( real )

[floral island]

casual flex xD

[vital ermine]

470 driver works?

[floral island]

9 minutes

[uncut fiber]

@noble shoal whats your problems with AIT?
Also one has to have old nvidia drivers to work it properly

[floral island]

i have to make it work -> that's what our engineers use because of a specific package they need to use

newer driver doesn't work well with that software

[noble shoal]

Only 1.600 Watts? That's the same as the hairdryer of my gf

[floral island]

our other server is in use -> that's the 80gb versions

[uncut fiber]

those thumbs could be nice they all creating some face. I forgot program that was capable to do it

[floral island]

fun times if i have to work in our datacenter and all the number crunching machine are at full power -> that's a pretty hot backside

[noble shoal]

My drivers are at the most modern ancient version for AI usage.

[floral island]

yeah, didn't AIT fail if you used the newest?

[uncut fiber]

AndreaMosaic i think

[noble shoal]

Not going to test updating

[floral island]

works on 531, i'm staying there

[uncut fiber]

531 here

[floral island]

i heard 537 has a lot of SD issues?

[uncut fiber]

and enable to disable, who invented priory to enabled... show me him/her!!! 😠

[noble shoal]

Fizzle Dorf did

[uncut fiber]

where he is!!! 🙂 cost me for sure most troubles restarts of comfui.

[vital ermine]

anyone know which node is doing this so I can delete it?

type in any test box and that pops up

[uncut fiber]

@floral island
http://www.mazaika.com/grafika/shining-actors-30x33_big.jpg
this i mean, like old version of SD 😄

[floral island]

oh yeah, that's was a extension too 😮

[uncut fiber]

it was extension? i know it only like standalone portable app

[floral island]

nah, someone made an a1111 in the early days i think, when it was just scripts

[uncut fiber]

o.k.

[noble shoal]

Nice British Punk Band Seinfeld

[uncut fiber]

hot woman

[indigo carbon]

I was actually working on that yesterday. The solution was to have the CNET do low steps, then send to another Ksampler without CNET and using prompt from BLiP, and it would be close to normal AIT speeds

[indigo carbon]

@fierce hollow this is what I managed to pull off for CNET

[uncut fiber]

how to calculate max images for resolution 1024x1024 16M colors?
(1024x1024)^16000 000?

[indigo carbon]

Anyways, if y'all want a solution for keeping AIT's speed while making it flexible for more stuff; I had nothing to do with this, I just compiled the modules. I completely agree that we DO need an exLLaMa equivalent for diffusion, the main goal would be to make it flexible and its speed being way faster than AIT is just a bonus.

[uncut fiber]

4× 10^96329598

[fierce hollow]

I don't mean like the output being bad, I mean comfyui completely errors out when I try to load a controlnet when AIT nodes are loaded, like even if I don't use AIT in the workflow it will still show an error if there's a controlnet

[ionic dragon]

no, its (16,000,000)! / (1,048,576)!

[uncut fiber]

o.k. going to try to numerize it

[ionic dragon]

dude, you cant

[uncut fiber]

i can.... on wolfram, pitty i didnt make a copy paste

[ionic dragon]

what did you get?

thats gonna be a hella big number

[uncut fiber]

o.k. going there to copy paste it

[indigo carbon]

The workflow I shared uses AIT

[uncut fiber]

Power of 10 representation
10^(10^8.010552413613613)

[fierce hollow]

it's actually 256^3, not exact 16M

[uncut fiber]

yes just for simplicity.

i could as well say 1000000

[fierce hollow]

doesn't matter, some modules are probably just broken, if you compiled your own that's probably why it's working for you

there's already issues about in the repo anyway

that said the ait repo seems dead and hlky deleted their account (wtf)

[inland vessel]

someone that sees it? 👀

[uncut fiber]

bottom is better seen

top bit worse

[inland vessel]

yeah for the top i have to blurr my vision lol

[uncut fiber]

on phones it must be visible well

[supple knot]

You can see them when there small like thumbnail size

[uncut fiber]

yes as thumbnail it isnt surprise

[inland vessel]

niceee

i love this stuff haha

[quick lake]

Hey, how should a prompt be structured to make the results as realistic as possible with the new Stable Diffusion XL, which I have running locally on my PC? What are the settings I should consider for achieving the most realistic images? Are there any other tips I should take into account? Thank you.

[supple knot]

Nice work I cant get mine QR system to work, getting errors

[inland vessel]

how can u make an image black/white

[supple knot]

photo editing software

[uncut fiber]

it isnt black and white

just in A1111 how to say english, cutted?

for examle this way. It is called rmbg

[supple knot]

havent seen anyone do this is ComfyUI or A1111,I think only post editing in another software, like photoshop

[indigo carbon]

Yeah, I kinda replaced hlky for a while, but the ultimate solution must be to make an "exDiffusion"

[uncut fiber]

This is in A1111 remove background extension, widely used

[noble shoal]

Rembg is also available in Comfyui

[uncut fiber]

yes and miracle one

[supple knot]

here a video i found on it

https://www.youtube.com/watch?v=Ki_ZcF_u23I @uncut fiber

[uncut fiber]

nice astronauts on bikes 😄

[uncut fiber]

Robots or mechs? Because seems headless and i cant achieve it with robots @vital ermine ?

[vital ermine]

robot human mech army

[uncut fiber]

o.k. TY!

[uncut fiber]

if you are using A1111 you dont need b&w images. Only you need Controlnet with specific model and one sd model. That CN models should soon ™️ arive in sdxl

[vital ermine]

we are about to get CN training we can do in our homes

[vale eagle]

humanoid

[vital ermine]

that should be interesting

[noble shoal]

Highly interesting

[uncut fiber]

laser bullet 🙂

[vital ermine]

multi functional weapon

explosive round

[uncut fiber]

multifunctional at once 😄

[uncut gull]

Good morning! It's a brand new day.

[uncut gull]

4K isometric scifi bedroom game asset. However, note that there is a pillow on the desk, and this is not a chair. It is problematic for a workflow.

[noble shoal]

Good Morning

[uncut fiber]

can be some sort of collapsible chair to spare place. Nice placed pillow what can be more comfy at work 🙂

[indigo carbon]

it should work if you have SM80.. anyways, if you want a newer optimization, we'd all love that

WE. NEED. EXLLAMA FOR DIFFUSION.

[uncut gull]

Inpainted at great difficulty. Probably needs manual painting of a base.

[fierce hollow]

on windows a lot of the modules are just not working sadly, exllama for sd in the sense of custom kernels or something would be nice but doubt it will happen, there isn't really as much interest as there used to be, llms are much more popular

[fierce hollow]

wish ait was just supported in comfyui by default, that would fix the issue of hacky patches and broken modules

[indigo carbon]

that's the only way though.. AIT is close to the speed limit you can get to without making stuff like exLLaMa

[indigo carbon]

it was actually planned to. I remember @visual glade mentioning that somewhere, but he decided not to release it due to the node working somewhat consistently

[fierce hollow]

yeah I remember that too

[indigo carbon]

I'm already telling, I won't be making new code for compiling CNET or IPA- it's way too dynamic and I don't have time for that atm

so if we would have a solution like exLLaMa it would likely be compatible with everything diffusion related

[uncut fiber]

@uncut gull i think you did nice work. If in A1111 there is great for inpaint outpaint OpenOutpaint, we talked about it today btw.

[fierce hollow]

I'd compile it in wsl or something but cba when there's constantly some new random thing that makes ait break... like animatediffusion, better unload cause it broken as heck with ait

[indigo carbon]

yeah, every time we try to make AIT work with newer stuff we wish there would be something like exLLaMa.. I'm sure it will happen eventually

[fierce hollow]

better set up a fundraiser for turboderp if you want exllama for sd

[indigo carbon]

I also saw FizzleDorf being able to use AIT along with TRT and they got to over 220it/s, Idk how the fuck did they manage to do that, but it was possible

[indigo carbon]

also exLLaMa is about times 8 as fast as pure pytorch, so it will be even faster than AIT

tested exLLaMa like 2 days ago, autoGPTQ did 8T/s while exLLaMa2 did 62T/s

[fierce hollow]

I mean you're basing that on llm speed though, it might not translate any better than ait to sd

[indigo carbon]

nah, AIT is an equivalent of DeepSpeed, which isn't anywhere near exLLaMa

[fierce hollow]

last I tried deepspeed in wsl it didn't really change much of anything so no doubt there

[indigo carbon]

I actually tried that on my dual boot, it was 14T/s instead of 8T/s, but still a fraction of what exLLaMa does.. so the idea is probably to create something similar to exLLaMa but for diffusion

[uncut fiber]

@hardy cipher even this is not readable. Tested on net qr checker. Maybe something can read it???
And there are way more complicated.

[uncut gull]

Is it an extension?

[uncut fiber]

yes with its own GUI.

For A1111

i already posted image i did it 100% in it.
#✨｜sdxl message

[indigo carbon]

http://testovaciobrazek/

It scanned

[uncut fiber]

so great it is working, why apps show not?

[indigo carbon]

.. idk

[uncut fiber]

probably it counts with some standards? Some weaker cameras, some sort of contrast or such?
@indigo carbon thanks great!

[uncut fiber]

curious if anybody realize theme of this qr code 🙂

not message in it, but how it was created

[urban fjord]

I tested different web QR sites and some reads it and some does not. (The grassy one)

[uncut fiber]

yes i think there is probably some standard.

[urban fjord]

Is that with SDXL? I had expected more blending between the two.

[uncut fiber]

according to fence it is sdxl.

wanna make gold shining through crack in cave and dwarfs protecting their eyes blinded by this shining, but nobody is protecting their eyes... This seems to be already blinded poor dwarf.

something like this

[ionic gulch]

after talking with the devleloper of the ip adapter node, the upcoming version of my workflow will have much more ip adapter options. for example up to 5 images and "merge as batch" or "merge as chain".
difference between merge as batch and merge as chain (all other settings are the same):
https://i.imgur.com/OpGeHMx.png
https://i.imgur.com/7NwA5mO.png

[uncut fiber]

😄

[ionic gulch]

5 input images and those options make it much easier to play around with different input images - for example turning on/off single or multiple images:

[uncut fiber]

isnt 5 too much? Yes combination of them probably

[ionic gulch]

5 input images also allow it to use the ip adapter as instant "lora light":
https://www.reddit.com/r/StableDiffusion/comments/16sglm9/no_more_lora_training_only_6_images_needed/

(he did it with 6 images in that reddit post, but from my tests, 5 or 6 images doesn't make that much of a difference) if needed i could expand to 8 input images, but to keep the menu area "single screen" that would have been too much unless it really helps for some usecases.

you also reach limits of even 24 gb vram if you add too many input images.

[compact hatch]

loras can be tagged and can utilize regularization images to prevent overfitting

[fierce hollow]

ip adapter is some sort of controlnet? looks pretty neat

[upbeat summit]

at some point it will work exactly like that, but this is good for styles I guess. right now it won't be as versatile as a LoRA fine-tuning

[ionic gulch]

that's why i called it "lora light". it may help in some cases and is fun to play around with it.

[shy kelp]

"vae fix" on left lol

[crisp owl]

left has vae issues, at least my left I see

[shy kelp]

its got weird artifacting, noticed it on a few stuffs

[crisp owl]

yeah that's why I always use manual vae node

1.0 vae is broken, and a lot of checkpoints still run with it

[uncut fiber]

this robot seems truly made in japan. Nice wheel chair 🙂

[shy kelp]

well idk where it came from but this was what it was

[uncut fiber]

0.9 is o.k. then?

[crisp owl]

0.9 is good

[uncut fiber]

thank you

[fierce hollow]

https://huggingface.co/madebyollin/sdxl-vae-fp16-fix
this vae fixes everything including fp16 training

[shy kelp]

this is what i am running now

[crisp owl]

https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/vae/diffusion_pytorch_model.safetensors

This is what I use, from SAI, I can confirm it works 100%

[uncut fiber]

did somebody make stats which sampler is better for text? Some accuracy or is it totall random 1:1?

[hardy cipher]

what sort of secrets have been revealed to JPS?

[ionic gulch]

the new workflow is online:
https://github.com/JPS-GER/JPS-ComfyUI-Workflows

[ionic gulch]

for me the most important info was that chaining and batching the images gives different results. combined with the "lora light" post on reddit, i've added quite a lot of options to the ip adapter part. we also might get batching with different weights if possible with an update of the ip adapter node.

[hardy cipher]

really curious what new things you have going on. I've been working with 3 ipadapter inputs and it worked, but just kind of winged it

[random zenith]

has SDXL already caught up to 1.5 for anime or not yet? Should I wait more before switching?

[hardy cipher]

no idea

[ionic gulch]

you can see most of the new options if you look at the new menu node:

[hardy cipher]

it lets you make eyes with skulls though

[shy kelp]

what do you mean caught up?

[ionic gulch]

some options are just "qol" - like using the ipa1 weight and noise for all input images, so you don't have to change 2x5 values

[indigo carbon]

is this good use of CNET right here?

[hardy cipher]

gotcha. I'll look at what you have going on. I've been saying they make LoRAs obsolete for a while. sort of sarcastically, but then they do serve a similar purpose for me

[fierce hollow]

nah, there aren't any good anime finetunes, only some claiming to be, wdxl seems dead, anime loras don't look too good trained on base

[shy kelp]

Where can i ask a question about the software

[random zenith]

I see , thanks

[ionic gulch]

combining up to 5 images with different weights, two different merge modes and the option to use FreeU gives so many different results for 5 input images, it will take some time to figure out my personal favorite settings.

[fierce hollow]

that lora light thing could possibly work for anime with sdxl, not sure

training 2 text encoders, base and refiner is just too much of a pain otherwise

[random zenith]

I hope a good finetune comes up eventually

[shy kelp]

[fierce hollow]

the only good anime finetune came from nai for 1.4/5 so not sure we'll see anything like that for sdxl

it kinda works, sure, but then you can do better with normal sd with half the vram

[indigo carbon]

and they where leaked xD

[fierce hollow]

yup, wd-whatever-version-they're at now is still worse than nai...

[hardy cipher]

1.5 might be good with anime, but can it make this?

[random zenith]

I only use SD for anime so I guess I'll have to be patient

[indigo carbon]

yeah, but NAI has leaks somewhere. when 1.5 made noise a leak of their anime version was thrown around the internet like grandma's 1 day old salad

[shy kelp]

hmm

[indigo carbon]

I think some people on 4chan hacked them or something? because they NEVER released anything for everyone to use

[hardy cipher]

yolo

[fierce hollow]

umm yeah, nai didn't release their model of their own volition, at least not as far as anybody is privy to know, it was supposedly leaked from some repo (by some employee? idk, it's old news)

[shy kelp]

Where can i find the terms and conditions for SDXL

[hardy cipher]

are you trying to do crimes with it?

[indigo carbon]

I don't think there are any really. it doesn't even need access to the internet after you have the weights

[shy kelp]

No you just wondering if I own what I make with it

[ionic gulch]

tried AC in front of a church with 3 input images. have to get some more face images of her to make it work better:

[fierce hollow]

https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/LICENSE.md

[hardy cipher]

ahh. https://stability.ai/terms-of-use

[shy kelp]

i mean, if you generate images of like 1:1 pokemon or mario and try to do a bunch of commercial stuff with it and those companies find out, you're gonna hear from their lawyers

[fierce hollow]

website tos is not the model tos, just check the repo

[hardy cipher]

gotcha

[indigo carbon]

that license is talking about multimodal generative models. SDXL isn't multimodal (yet)

does this mean SAI is planning to make a multimodal version?

[fierce hollow]

it's just standard openrail license wording afaik

[indigo carbon]

gotcha. for a second I thought this confirms if we'll get a BLiP diffusion SDXL version eventually haha

[shy kelp]

So if I use it to generate nature or like non own characters or things that got a trademark for them. And then I own the images it generates or not

[true wing]

I'd be surprised if stability isn't working on something multimodal at this point. That seems to be the direction all these companies are headed and there are already open source text, audio, and (rough) video models/techniques.

[shy kelp]

dont see why not

[fierce hollow]

sd copyright is still gray zone afaik

[shy kelp]

ai copyright in general is gray zone tbh

Well I think I stick to gimp and krita and Inkscape

[fierce hollow]

there were some cases recently where judges ruled ai art can't be copyrighted, but there was another where copyright was given to a game or something using ai as part of it

hard to say how it really works in the end, I guess transformative work counts

if something is created completely by AI, then it is not copyrightable; if it is created by a human with the support of AI, then it could potentially be copyrightable.

or so it said

[shy kelp]

very cool there are judges who make rulings on tech stuff where they themselves don't know how to forward a simple email and get scammed by a fake irs phone call

[fierce hollow]

iirc steam removed some games that used ai art recently too, while epic's tim sweeney said they're welcome on the epic store

[hardy cipher]

yeah, it's a common problem throughout oall of politics really

from the bottom to the top

and the legal system

[shy kelp]

oh yeah, gabe definitely doesn't want to touch ai assets in games so they just outright ban ai usage on steam

[hardy cipher]

people get in those positions, city council, judge, senator, etc, and they are suddenly, or over time, convinced they're experts on everything. they always know best

I like that the average US senator is at the age where a lot of people are entering nursing homes

[pure crystal]

burger lovers

[hardy cipher]

they're face melding

[uncut fiber]

robots have nice guitars

[indigo carbon]

what CNET does this? I found this somewhere but they didn't explain what CNET was able to do this

[uncut fiber]

qrmonster i believe for SD

[indigo carbon]

not XL?

[uncut fiber]

not xl now but author is planing it

[crisp owl]

don't think anything sdxl does that easily yet

[hardy cipher]

the gay cats one blew my mind the most

[indigo carbon]

wait, so wtf is this? https://huggingface.co/Nacholmo/controlnet-qr-pattern-sdxl

[noble shoal]

Unfinished

[uncut fiber]

o.k. somebody said it is easy to train it

[indigo carbon]

wait, so this isn't possible for SDXL yet?

[noble shoal]

The results are not good. I have tried it for an hour or two, but it's really not usable imho

[true wing]

I think the thing isn't so much that Gabe is against AI as it is they don't want to deal with the legal head ache until there's a definitive resolution. Which is probably going to take at least a year or two with all the court cases. Epic on the other hand doesn't care because it 1.gives them an advantage and 2.they don't deal with the same volume of games on their platform so the risk is worth the benefit if more people use the Epic store.

[indigo carbon]

just looked into it, isn't QRmonster made for just QR codes? those images did words

[uncut fiber]

it is made by text imo

mmnt

[noble shoal]

I think it's more about black and white values

[hardy cipher]

it started with qr codes I believe, but the same approach can be used for other things

[indigo carbon]

also, does that need any preprocessing? I can't see any nodes made for QRmonster

[noble shoal]

Thresholding

[uncut fiber]

no it doesnt need preprocessing imo. Maybe switch b with w

[noble shoal]

The image should be black and white. Not gray-scale

[indigo carbon]

so canny preprocess with QRmonster CNET makes those images?

[noble shoal]

I think you need thicker lines. I would recommend an image threshold node

You want to get the shadow values as black

[hardy cipher]

well I don't mean to be pedantic, sirs, but I do believe canny uses thresholding

[uncut fiber]

o.k. can be subjective. I told it make it out of mouses, and i it does from computer mouses. AI surprised me

[noble shoal]

But you need to separate the brightness values instead emphasizing the edges

[hardy cipher]

but I realize it is different. anyway, just use this https://docs.opencv.org/4.x/d7/d4d/tutorial_py_thresholding.html

[noble shoal]

There's already a node 😉

[hardy cipher]

which one?

[noble shoal]

I have no idea. I am at work, but I saw it this morning

[hardy cipher]

gotcha

[uncut fiber]

it is fun to experiment. Hiding color images is posible as well

how is called this projection? Ultrawide?
btw for those going to try it denoise at 1 is o.k.

[pure crystal]

as long as it's not a spiral

[crisp owl]

skull earth theory

[uncut fiber]

complicated 🙂

[shy kelp]

just getting into sd, is there an xl model people can suggest or everyone still using the default? 🤔

[uncut fiber]

There is too much versions, nightvision, copax. It is all subjective. Some ppl using still default

[crisp owl]

RealVis, Juggernaut, Nightvision (protovision), ZavyChroma, CopaxVivid

Those are some of my most used ones

[shy kelp]

thanks. nothing that seems to be head and shoulders above the default though huh?

[crisp owl]

I don't use default

[shy kelp]

hrm, let me be more speicifc maybe, any xl model you might suggest for photo reals or do they all have their plusses and quirks

[crisp owl]

I mostly do photorealistic gens, basically the list I did kinda goes in order of photorealism preference for me

the more "artsy" I want tends to push more to the right side. Currently I'm using CopaxVivid for those skull ones

[uncut fiber]

also depends if you want or not nsfw.

[uncut fiber]

or anime 🙂

[indigo carbon]

getting this with qr monster on 1.5

[crisp owl]

just one more drink of earth

[uncut fiber]

with 1.5 model? I am using A1111 for it.

[pure crystal]

no more sips of earth before dinner

[uncut fiber]

probably AIT turn off?

[noble shoal]

Does it work if you just feed an image, like black text on white bg, into the cnet?

[crisp owl]

"forget this, I'm out"

[indigo carbon]

no, always gives that error

[inland vessel]

sai?

[uncut fiber]

yes

[inland vessel]

i love pasta ones

lol

[uncut fiber]

send nodes!

[noble shoal]

Ok, then I can't help atm. Maybe you can drag drop one of those images

[inland vessel]

also funny how it made a hand fork

[hardy cipher]

tuple out of range, bud

[uncut fiber]

no text here 🙂 nothing here, just skeletons hands in hands

[hardy cipher]

skeleton island

[uncut fiber]

what projection is this

[hardy cipher]

skeleton waterfall

[hardy cipher]

I forget how I made that. ipadaptor and controlnet I think

[uncut fiber]

np 🙂

[hardy cipher]

I could figure it out

if you squint hard enough they're eyes

[uncut fiber]

yes i saw it

interesting

[hardy cipher]

eye made a few of them

I've just been going through stuff I made recently

sometimes don't even end up looking at it initially, or just skim over it

kitties

[uncut fiber]

same

[ionic gulch]

copying a style works quite well. better than specific faces:

[shy kelp]

is it possible to make XL embeddings in the automatic ui? or only 1.5?

[uncut fiber]

Mondrian artist is great for playing with controlnet.

[hardy cipher]

you need this

I really just want to know where I'd get meat carpet like that

[uncut fiber]

@indigo carbon have you solved it? A1111 is easiest for now it seems

i like this lora

[upbeat summit]

it should be out soon

[indigo carbon]

yeah, currently restoring my old A1111 install

[uncut fiber]

i am quitting smoking and having hard times... Smoked long time without nicotine long time, but still difficult not use it again.

[upbeat summit]

you can do it! keep playing the latent space dopamine machine instead.

[uncut fiber]

🙂

thank you

[upbeat summit]

the fight is to avoid the next one. the hardest are the ones that were like a "ritual" - after eating, coffee, when going out. this will get better.

[hardy cipher]

indeed

not sure how to feel about this one

[crisp owl]

with addictions, the hardest is the "boredom" triggers. SD is great to alleviate that lol

[hardy cipher]

I quit drinking largely out of spite. don't let anyone tell you that can't motivate you to do better for yourself

[hardy cipher]

but smoking would probably be harder, I do understand that

[crisp owl]

They're all hard in their own right, think everyone has something they wish they could kick one way or another.
I know I do

[uncut fiber]

yes but without nicotine i am quite long time already. and still 🙂

[noble shoal]

So you vaped?

[uncut fiber]

yes last time.

i simple not used nicotine later. I used to smoke 40 cigarettes cca

[upbeat summit]

yeah well getting rid of nicotine is already a pretty large step

[noble shoal]

True that.

[hardy cipher]

I do believe vaping is a whole lot better than smoking. my mom smoked for decades and vaping got her off the cigarettes. might not be as healthy as not doing anything like that, but certainly better than cigarettes

[noble shoal]

I vape for 4 years now. I smoked a lot, really a lot

I can buy a 3060 each month for what I safe