✨｜sdxl

[hardy cipher]

smoking is so vilified in some places now. in california, good lord. the advertisements basically conditioning people to treat smokers like monsters, lol

[upbeat summit]

avoiding dozens of burning toxic metals in your lungs is also a good thing

[uncut fiber]

i seems to be for some thing that is in vaping thing.
And in USA they basicaly pushed ppl to smoke in 60-80s

[hardy cipher]

I get that it's bad and all, but need to focus on the root of the problem

[upbeat summit]

the problem is addiction... but I just want to run this batch of 30 images before going to sleep

[hardy cipher]

tobacco lobbyists and the whole structure that pushed it onto people

well I just leave batches running when I sleep

easy peasy

wake up to a stack of new images

[uncut fiber]

got idea i will made picture with stop smoking text hard to read it 🙂 so it will work dont know how in english

[upbeat summit]

same - I just want to see how the prompt does 😄

[hardy cipher]

when donald trump made a baby with chris christie

[uncut fiber]

he has also czech emigrant for wife.

[hardy cipher]

masslevel: have you used the dynamic thresholding much? trying to understand what a couple of the values do. interpolate_phi for example. I see it there, and I mess with it some. but no idea what it's actually doing. and that code is some tricky stuff to look at

[compact hatch]

The worst thing to happen to anime models. It has severely stunted them.

[uncut fiber]

Missis owl and mister owl. BUBO BUBO

[upbeat summit]

I think we talked about it a couple of weeks ago. I used it for some experiments in a1111 with SD2 and I really kept to mimic_scale, threshold_percentile, mimic_mode and cfg mode. I suppose interpolate_phi will smooth out some of the stepping between the mimic and cfg, but that's only speculation on my part.

I looked into the code but - that's way over my head 😄

[uncut fiber]

@indigo carbon how it looks? i only used invert preprocessor, looking forward for your tests!

[hardy cipher]

well I'd talked to you about the extension in the past, but not the new node. there are a couple extra things slipped in there which is great, but can't figure out what they are

[upbeat summit]

I haven't really tried it out in ComfyUI to be honest. I tried to find out what interpolation means, but I didn't find anything. I will do some experiments myself later - maybe we find a pattern 😄

[hardy cipher]

I'll normally turn something all the way one direction the the other to see what happens, the npull back from those extremes

[uncut fiber]

interpolation of what?

[upbeat summit]

same 😄

[uncut fiber]

i know only video interpolation Riffe or Rife <- which is best!

[hardy cipher]

also not sure about that variability measure. I tend to just switch it back and forth every so often

[uncut fiber]

i made gorilas as well

i add in prompt out of gorillas with room in background. <- for example

[uncut fiber]

man with wooden mask

[upbeat summit]

you could ping mcmonkey 😉

[uncut fiber]

good start

[hardy cipher]

yeah. he actually let me know that he'd made that node

I'm sure he'd have a proper answer since he did make it and all

[uncut fiber]

or garden wall with flowers and those i add in this image

[uncut fiber]

you can give this in competition i think or it already ended?

[hardy cipher]

hmm, me? the sticker thing?

[uncut fiber]

no cow submission

some weird invention was theme

[hardy cipher]

cows are cool

[uncut fiber]

it ended yesterday 😦

[hardy cipher]

cursed

[noble shoal]

Sticker is still happening. I made a PoP rainbow goblin 😅

[uncut fiber]

@indigo carbon try this 🙂

i saw it you have nice things in stickers. @noble shoal

[indigo carbon]

this chat is SDXL, that's 1.5

[uncut fiber]

o.k. moving back

[noble shoal]

Thank you. The problem is that most people outside of this chat will not understand the references.

[uncut fiber]

yes but it looks nice, also not expecting avalanche of voters.

1.5 has not place to post images. Good night!

[noble shoal]

Good night. And keep it on with not vaping/ Smoking 💪

[noble shoal]

Oh my. What is he touching there?

[hardy cipher]

I'm be honest, I never looked at it that closely
But it's clear that is a wood carving of some kind

[pure crystal]

it's a peen

[noble shoal]

You might be right with some kind of wood.

[noble shoal]

From the texture it could also be a poon

[hardy cipher]

Get your minds out of the gutter, everyone. I'm offended

[noble shoal]

Appreciate our art critic!!!!!!

[shy kelp]

the fine lines on the turtleneck and shirt

[cyan crown]

https://youtu.be/bVuGvjvqO2c

[shy kelp]

half-closed eyes, bloodshot eyes these racoons

[pure crystal]

racoon dj wearing cap, dj equipment, light show, ( POV:1.3),(adult, night time, horror), 1997, (extremely detailed depth), style of R Crumb, highly detailed, style of comic book cover illustration, inspired by William Brodie

[shy kelp]

hmmm

[shy kelp]

hell yeah

[shy kelp]

(i am currently messing with a model merge atm)

[shy kelp]

[shy kelp]

reminds me of an unintended generation i got the other day

[hardy cipher]

Trash pandas and skulls. My kind of party

[shy kelp]

that and the red station wagon goes hard

[pure crystal]

lol 1983 chevy suburban

[shy kelp]

i can't stop putting the word three-quarter view in my prompts lol

[crisp owl]

It's easy not to!

[crisp owl]

Keeping his bones for himself

[hardy cipher]

Cfg 50

Need that keyboard

[indigo carbon]

who needs CNET when SDXL can just write on its own

[lusty wolf]

RedHot!

[wary flint]

Cyberpunk Elephant

[lusty wolf]

Been playing with Comfy much too long...

[delicate kelp]

Some Tilt Shift Miniaturization (Lora I'm working on) to brighten your day

[hardy cipher]

some luck munchkins to brighten your day

[shy kelp]

hard to find a balance in this merge but i feel like it's around here

[uncut gull]

HOLD UP!!! Yesterday someone here had a moving skeleton? Was that made in SDXL? Is there a way to use SDXL for video???

[shy kelp]

probably, think there is deformer? idk, i am probably wrong

[noble shoal]

Or was it animatediff?

[uncut gull]

It looked exactly like animatediff. The repo has no word on retraining for sdxl though. https://github.com/continue-revolution/sd-webui-animatediff/issues/82

Is there a way to search for the post? Probably not...

How was this made???!!! Was it SDXL?

[shy kelp]

yeah that does look like animatediff

[hardy cipher]

https://tenor.com/view/always-has-been-among-us-astronaut-space-betrayal-gif-23836476

[hardy cipher]

you ready for this, piscabo?

time to kick it up a notch

[nimble heart]

Why is there "prompt" metadata if the workflow meta should have everything?

[hardy cipher]

perusing metadata on your phone?

[nimble heart]

why not? I just found out my gallery app can view the png text chunks

mostly useful for a1111 images cause comfyui meta isnt super readable without deserializing the data

[hardy cipher]

that's cool. and I don't have an answer to your question. I wish I did, but I suffer from iq deficiency

what's it going to make?

[nimble heart]

are you remixing my screenshot

[hardy cipher]

yeah

[noble shoal]

More metadata

[hardy cipher]

it isn't that good though. sorry

[nimble heart]

what if it generates the perfect workflow

[hardy cipher]

look at those beauties

[nimble heart]

wonder if you could feed like the top 50 workflows from civit.ai into chatgpt and have it produce the ideal workflow

[hardy cipher]

maybe the api

the web version has the memory of a fruit fly

[nimble heart]

ah. never used

web only like 4k context or something?

[hardy cipher]

it's maddening to try doing anything complex

[nimble heart]

or daresay even less

[hardy cipher]

I heard 4k, maybe 8k. but good luck with something that uses a lot of random characters and things

because it's like a vacuum and it sucks up literally everything

and then pushes out something else

[nimble heart]

4k seems fine enough in my local install

maybe my expectations are low cause I come from AiDungeon land lol

[hardy cipher]

problem is it doesn't tell you it's forgetting things. it doesn't know. so it'll just go on like things you said were never said. things it said were never said

[uncut gull]

It just told me to use A1111. (I'm joking LOL. But that is the point of everything other than ComfyUI. Trying to be "the" perfect way to use SD, instead of customizable.)

[nimble heart]

does it have a memory store? so you can push key details into permanent context?

[nimble heart]

who told you that?

[uncut gull]

ChatGPT, when I asked it for the perfect ComfyUI workflow. 😛

[nimble heart]

Ah

I still have a1111 installed I just dont use it for 90% of stuff

[hardy cipher]

how did I miss this beauty?

[indigo carbon]

Bing GPT-4 has access to the internet. that might tell you what the ultimate workflow is

[steady grove]

these language models are idiots and will just make shit up on the fly

[nimble heart]

maybe once my gpu can use the pytorch attention I'll try it out again but for now comfyui's sub quad is by far the fastest attention for AMD gpus

[steady grove]

don't believe their lies

[hardy cipher]

bing gets really weird when you start asking too much about image rendering stuff

[steady grove]

they don't actually have any understanding of better. they just predict the next letter and make it sound believeable. They're experts at "truthiness"

[hardy cipher]

I had it end a conversation when I asked some stable diffusion related questions

it thought it was stable diffusion. said it couldn't tell me it's secrets

[noble shoal]

There was a time it ended the convo when you mentioned "prompt"

[uncut gull]

It's been pretty hard for me to get SDXL to render many different characters in one prompt.

[hardy cipher]

it ends a lot of conversations

[steady grove]

its not thinking. ugh. James Cameron is awesome but he sure ruined everyone's ideas about how AI is supposed to work. Terminator is his Jaws.

[hardy cipher]

sometimes I'll start a new conversation and call it a psychopath. then it ends that one too. nothing gained

[uncut gull]

Oh! Do you think AIs don't think? I have some papers for you!

[hardy cipher]

they are very adept at predicting what comes next

it's a form of thinking I guess. but it's hard to define where that line is

[steady grove]

They execute and regurgitate encoded routines from training.

papers are a dime a dozen. Most are garbage. This is the peer review process and how it works.

[hardy cipher]

needs more elephant

[uncut gull]

Ah hah! Here, OthelloGPT. When trained on the moves from Othello, the GPT learned to predict the next move. No surprise there?
How did it do this? Using latent saliency probes, researchers from MIT and Harvard looked inside the layer activations, and found this: It had developed the concept of an 8x8 board, black and white tiles, the rules of flipping states, and the rules of legal and illegal moves. 🙂
Why? Because that's how you predict the next move in Othello. By learning the rules. No other way.
https://arxiv.org/abs/2210.13382

[crisp owl]

it's not thinking.
When there is no input, there is nothing they are doing.

[uncut gull]

When it is not thinking, it is not thinking.
What about when there is an input? Then it thinks. 🙂

[crisp owl]

So that's not thinking

[hardy cipher]

bard thinks about how it's going to lie to me the next time I make the mistake of asking it about something

[crisp owl]

It's using algorithm's to predict what's next based off weights and values

[uncut gull]

First of all, I doubt you know what you mean by those words.
But let me ask you this: What's the difference between the way you think and the way an LLM thinks?

[crisp owl]

I'll stop there if you're gonna throw that at me, thanks

[uncut gull]

LOL well I have to go anyway. But LLMs build internal world models governed by rules and have analogs to theory-of-mind neurons in human brains. Stable Diffusion also doesn't copy-paste images from its memory, it builds out a 3D scene and then starts slowly refining the figures, adding color and detail.
The difference between how you think and how an AI thinks is: We can look inside an AI while it's thinking. We can figure out exactly how it does what it does. And there's no surface-level statistical memorization about it. These things think . 🙂

[hardy cipher]

stable diffusion looks at noise and denoises it right?

[uncut gull]

Indeed! The question is: How does it do it?
Here's a metaphor, if research doesn't convince you:
The only way to predict how bright a person's skin will be in a scene is to figure out the rules of light.
The only way to predict what a person would say in a scene is to figure out the rules of how people talk.
Figuring out those rules is what AI does.
(As for what humans do, that is still very much up for debate.)

[hardy cipher]

I know how it does it. it's the training in reverse

it was trained to isolate the noise in images

[steady grove]

Can a machine think? I don't think conciousness is magical. It exists therefore it can be engineered. Are LLMs thinking? No. They're orders of magnitudes smaller networks than any kind of sentient brain

[hardy cipher]

in ever increasing amounts

[uncut gull]

You don't know the difference between consciousness and sentience, and you don't have a definition of thinking. Not worth arguing. We might as well be talking different languages here. And I'm going to be late for my writers circle. TTYL

[steady grove]

they're also structured in such a fundamentally different way, that they are not analogous to "thinking" and should be described much differently

[hardy cipher]

I also think human vanity might get in the way of us entertaining the idea of a machine being conscious. but at this point I don't think it's really an issue as we're still in the infancy of the tech

[steady grove]

nobody can define consciousness or sentience entirely so , cool catch 22 you've found yourself there.

"i'm right because you can't tell me the exact differences"

[steady grove]

yeah. There are implications and paths in the future. We're just not there

[hardy cipher]

elephant hoods

dudes in the top left with their minitrunks

[wary flint]

😅 I'm glad you liked it

[hardy cipher]

used them to make this

brewing something else now

[wary flint]

Thats awesome! I don't use comfy, I was wondering if it was possible to use multiple images to generate new images

[hardy cipher]

it's my own frankensteined setup

[shy kelp]

organize this, now

[hardy cipher]

it is organized, what do you mean?

[shy kelp]

[hardy cipher]

wat?

[shy kelp]

lol big ass head

[hardy cipher]

honestly it's a bit hard to organize at the moment since I keep adding and removing junk

[shy kelp]

wonder how well this new merge can handle horror, so far it seems good on other fronts (these are supposed to be a mallet and hammer though )

[lusty wolf]

Add this...

[shy kelp]

monster made of lovecraftian junji ito flesh

oh 😳

[shy kelp]

lol oh god

[hardy cipher]

surprise!

[shy kelp]

NESSIE NOOOO

[hardy cipher]

I did pick a rather odd style for that though. could make it monsters if you want

[glad grove]

make it a sexy mushroom/dinosaur girl

[hardy cipher]

you know, that's pretty accurate

[shy kelp]

me fr lol

[icy brook]

Just released Aether Glitch.

https://civitai.com/models/151542/aether-glitch-lora-for-sdxl

[shy kelp]

vhs like, love it

a glass dodecahedron bomb, glitter, gold flakes, bright emissive inner core, wires, gears, tubes, cables, metal, symmetrical, crystal, cyberpunk, steampunk, fantasy

[hardy cipher]

my goodness. these are going to be something

imagine walking outside and those things were just standing there staring at you

I feel like these concepts have never been combined before

we're breaking new ground here

[wary flint]

lol the lower right one looks pretty cool. I'd put that on my desk

[hardy cipher]

this styler probably has thousands of options. good lord

[shy kelp]

hmm
gross

[hardy cipher]

living meat tree

alright, doing this one in the style of "bangkok"

well alright

[shy kelp]

nap time

[noble shoal]

I am interested. What styler are you talking about 🤔

[hardy cipher]

one second. I'll find a link to the github

https://github.com/wolfden/ComfyUi_PromptStylers

there's a ridiculous number of styles between those

[noble shoal]

Thanks. Going to borrow some of those json files

[hardy cipher]

just look at the milehigh one on it's on. haven't counted but there have to be at least a couple thousand styles on the list? at least it seems like it

lol, wtf

[uncut fiber]

can you share workflow or must be build itself, several hours? 😄

[hardy cipher]

me?

[uncut fiber]

got it already

[hardy cipher]

oh, well then why'd you ask?

[noble shoal]

Holy shit. 1521 Styles only in that json.

[hardy cipher]

there's a conjoined twin style

[noble shoal]

Looking at the amount of styles and prompts, I think there is even worse stuff deep down.

[uncut fiber]

i didnt know it is in png 🙂

[hardy cipher]

hmm. I wonder which settings you got

it's quite fluid

I saw something worse than the hate crime one. but discord won't let me use the words

[noble shoal]

@hardy cipher convert the Style to an input, connect it to a primitive and set it to randomize

[hardy cipher]

hmm. that's a good idea

fyi, don't use more than at most a couple thematic stylers or they entirely take over whatever you're trying to do

unless you do something like lower their weight and remerge with the original prompt

some of those stylers are more for structure of the image though, so those might not be as bad

you should see how many styles are in "all"

wait, that's just artists. that name is a lie

[noble shoal]

You got scammed. Open an issue and complain about the name. But how many lines does that file has? It takes ages to load on my phone

[hardy cipher]

jeez

[uncut fiber]

are you using this WF?

[hardy cipher]

probably as many as that other one. I don't even know

que? what is that?

[uncut fiber]

if anybody has some lighter i would love if somebody share it. Thank you.

it is from image on that page, only one rendered image there

[hardy cipher]

the styler page?

[uncut fiber]

https://github.com/wolfden/ComfyUi_PromptStylers/blob/main/examples/workflow.png
yes

[hardy cipher]

well I don't know anything about the workflow from there. just used the nodes

[uncut fiber]

https://github.com/wolfden/ComfyUi_PromptStylers/blob/main/examples/4.png

from here i have it

[noble shoal]

Just build this WF : https://github.com/wolfden/ComfyUi_PromptStylers/blob/main/examples/sdxl_prompt_styler.png

[uncut fiber]

o.k. going to build it

i know

[hardy cipher]

build what?

[uncut fiber]

wf

[noble shoal]

The workflow

[hardy cipher]

my workflow?

I wish you the best of luck, my friend

[noble shoal]

I think we are talking about a simple styler workflow 🤣

But yes. I wish everyone good luck building yours.

The GitHub links were direct links to pictures on the GitHub repo. Confusing that discord makes this big preview

[crisp owl]

you can just remove the embed

[noble shoal]

Also on mobile?

[crisp owl]

mmm, maybe long-click?

[hardy cipher]

my flow actually isn't that complicated really

not compared to some of the beasts that I've seen

[noble shoal]

I had some serious lagging viewing some other workflows. And I was not even generating.

[uncut fiber]

how i add input to sdxlpromptstyler? please

[crisp owl]

hmm nope just checked in my personal discord server, can't seem to remove the embed on mobile lol

[hardy cipher]

the fun and games don't stop here

hmm, I should bump up those steps now that I have it a bit tighter

[noble shoal]

Couldn't find a solution either.

[hardy cipher]

put it on random like you suggested

[noble shoal]

Add a primitive node to each of the inputs.

[uncut fiber]

i have not inputs. i am lost 🙂

[hardy cipher]

neon cyberpunk

why do they keep having food? lol

[noble shoal]

Oh, you have 2 textboxes in the node itself. Pos and neg

[uncut fiber]

got it

[noble shoal]

But you could right click on the node and convert those to inputs

Let's say you have another source for your text. Then that option comes in handy

[half ivy]

These are dope 🔥

[uncut fiber]

@noble shoal thank you just second before you told me i realized 😄 Thank you VM

[noble shoal]

Is that the meta logo?

[hardy cipher]

ha. nah, just mixed some images in a weird way. some styles definitely take over a lot moer than others

[noble shoal]

No problem. You're welcome

[uncut fiber]

not compatible with AIT or low memory

[noble shoal]

The styler?

[hardy cipher]

styler shouldn't take much vram at all

[uncut fiber]

i get error using AIT

[hardy cipher]

shouldn't be from that right?

[noble shoal]

Pretty normal lately

[uncut fiber]

out of VRAM supposing not compatible

[hardy cipher]

seems like text input should not do that

it's not computing anything

[uncut fiber]

Who asked for embedings error?

i believe it is primitive node

[noble shoal]

Did you set it to disable?

[uncut fiber]

yes i did

[hardy cipher]

yeah, I feel like using ait is a nonstop battle

[noble shoal]

AIT lost. I successfully removed it

[hardy cipher]

I installed it 3 times and then uninstalled it 3 times

doges won this round

[uncut fiber]

so purpose is for example when photo style it can show photo from 16th century?

[noble shoal]

Is it still randomising?

[hardy cipher]

yeah. in that mile high styler and then a depth styler

[noble shoal]

Anything can happen

[hardy cipher]

I should do the same with model weights. really get wild with it

[uncut fiber]

what is log_prompt oh record it?

[hardy cipher]

prints it in the console

cyberdoges

[uncut fiber]

3rd definitely normal

[hardy cipher]

4th one is mirrored doge

left and right sides

[noble shoal]

Ok Girls and Guys. I need my beauty sleep. Good night.

[uncut fiber]

@noble shoal any request?

good night @noble shoal

[hardy cipher]

it just adds words to your prompt essentially

[noble shoal]

Thanks

[hardy cipher]

https://tenor.com/view/goodbye-goodbye-little-sprout-the-barbarian-and-the-troll-sad-so-long-gif-21352802

thicc

[uncut fiber]

paper thing. It is very bad typing letter from embedings it whisper it

[hardy cipher]

que?

[uncut fiber]

when i type in primitive node and type letter from emb..... it whisper me word embedding:

so somebody asked today what can caused some embeddings flood, and i think it is primitive node

[hardy cipher]

hmm.well I don't quite get what you're saying here

[glad grove]

🤷‍♂️

[hardy cipher]

hmm.well I don't quite get what you're saying here

[glad grove]

the J

[uncut gull]

The models are lopsided?

[hardy cipher]

are they putting everythi8ng on one sideo f the image?

lol, why are you typing into the primitive like that though?

good

that's what you get for making degenerate images

[uncut fiber]

where should i type it?

[hardy cipher]

I'm just kidding beepboop

anyway, it's drooping or is the entire image jacked?

[uncut gull]

LOL. Have you tried without hires fix? 1024*768 is not a resolution I have tried using. Try 1024x1024.

[hardy cipher]

bernix. you didn't need to make your text boxes inputs. you can undo that. the primitive is for the style so you can randomize that

there's no reason to add a primitive to type in text when you can already do it in the styler node

[uncut gull]

1024x1024. No hires fix. No negative prompt.

[hardy cipher]

it's not, that's why I said his name in the previous response

[uncut gull]

Only if you're getting something you don't want. Then specific negative that out, but it will always hurt quality if you do.

[hardy cipher]

well if I were you I'd try to simplify everything and see if it still does it

change model, se if it still does it

change resolution

etc

that will help you determine what the issue is

[uncut gull]

Yep. Always makes small faces more distorted, hands more distorted. Even simple negatives like "blur".

[hardy cipher]

and what were the results?

and is a1111 up to date?

this stuff isn't like a normal program that you can update once every month or two

[uncut fiber]

1024x768 should work

[hardy cipher]

if you go a few days it'll often just break on you

and negatives don't hurt image quality, lol

you're rendering the image at the wrong resolution

[uncut fiber]

how randomize? @hardy cipher

[hardy cipher]

it's not an opinion

negatives can hurt images, sure

[uncut gull]

Something is very wrong and I have no idea what.

[hardy cipher]

it's supposed to be 1 megapixel

[uncut gull]

They don't always hurt quality. But I don't pile them on needlessly.

[hardy cipher]

1024x768 is not 1 megapixel

ahh

well

[uncut gull]

There is something broken.

[hardy cipher]

have you tried multiple models then?

[uncut gull]

Try updating?

[uncut fiber]

but i tested it and it is working. Probably you are using non XL model?

[uncut gull]

Which model?

[hardy cipher]

he's using juggernaut

which is what I'm using

[uncut fiber]

jugernaut is as well not XL version

[uncut gull]

The image you posted is 2048, not 1024. You'll get duplication at that size.

[hardy cipher]

I could try to walk you through this, but it's too much outside noise

[uncut fiber]

o.k.

[hardy cipher]

bernix

change your style to input

[uncut gull]

Turn off hires and try again, please.

[hardy cipher]

create a primitive

[uncut fiber]

try 1344x768 or 1536x640

[hardy cipher]

connect to the style input

change to random

[uncut gull]

Try typing upside down while drinking a glass of water.

[hardy cipher]

can I see your settings again?

what you're using now

[uncut gull]

Prompt lopsided droop at 1024x768.

Yes.

[hardy cipher]

cursed plant

dude, it's not the neg

stop

negs might make it not look ideal

yes, they were wrong

[uncut gull]

I didn't think hires could mess up that bad at 0.2 denoise though.

[hardy cipher]

negs can make something look bad, yes

but not deformed like that

[uncut gull]

Exactly.

I have no idea. Try again with 0.1 denoise?

[hardy cipher]

how about try without hires?

that plant image is cursed btw

[uncut gull]

Are you sure? Maybe it's just misunderstood.

Hang on, testing now.

[hardy cipher]

man

[uncut gull]

0 steps.

[hardy cipher]

lol

[uncut gull]

Well, that was the first thing I tried, and it worked. 😛
Trying another option now.

[hardy cipher]

well anyway, I'll let you two do your thing

good luck

in the future try to render everything at 1 megapixel

[uncut gull]

Yeah, and then there are other weirdly specific dimensions that work.

0 steps doesn't look great honestly. I'm trying 50 now.

[hardy cipher]

https://stability.ai/sdxl-aws-documentation

scroll down to resolutions

[uncut fiber]

0 is same as render steps

[hardy cipher]

about 75 percent of the way down

that is the list of resolutions that will work well

[uncut fiber]

i know they suggested at least 20 steps

[hardy cipher]

that's the official stable diffusion information from their site

no problem

[uncut gull]

That makes more sense!

[hardy cipher]

basically you want to get as close to 1024x1024 pixels as possible without going over, kind of like the price si right

why are you guys worrying about the high res steps? lol

[uncut fiber]

anything looks bad here?

[hardy cipher]

it has nothing to do with things being off

[hardy cipher]

looks like sweet treats

[uncut fiber]

768x768 🙂

[hardy cipher]

768x768 can often work

[uncut gull]

0 and 50 both worked perfectly for me. Trying 15 now to see if I replicate your failure.

[hardy cipher]

but issue is with things like faces, it's easier to see when it goes wrong

bro, it's not the steps

good lord

[uncut fiber]

o.k.

[hardy cipher]

bernix, not you, lol

[uncut gull]

At all your exact settings, I do not get those results. Try restarting A1111. Try updating it.

[hardy cipher]

it is the resolution

or the model perhaps

[uncut gull]

Hires 2048*2048, with 15 steps at 0.2 denoise.

[hardy cipher]

looong face

it's not the hires or the steps

[uncut gull]

Restart. And maybe update.

Not running on your local PC?

[uncut fiber]

there isnt portrait in styler

[hardy cipher]

are you positive it's the xl version of juggernaut though?

[uncut gull]

Okay... No further advice, sorry. 😕 Maybe they have a bugged out version of A1111 that's running the hires at the wrong denoise.
Try denoise 0.8 instead? Maybe it's running backward??? And 0.8 is actually 0.2?

LOL that's a totally stupid guess. I have no idea.

[uncut fiber]

probably i have bad vae

[uncut gull]

Yeah, sorry.

[uncut fiber]

i am using jugernaut as well but in comfy and not any issues.

[hardy cipher]

it's too bad they strip all the metadata from your image

I could normally load it and see what's going on

[pure crystal]

Stable Diffusion XL Resolutions
For the best results, it is recommended to generate images with Stable Diffusion XL using the following image resolutions and ratios:

1024 x 1024 (1:1 Square)
1152 x 896 (9:7)
896 x 1152 (7:9)
1216 x 832 (19:13)
832 x 1216 (13:19)
1344 x 768 (7:4 Horizontal)
768 x 1344 (4:7 Vertical)
1536 x 640 (12:5 Horizontal)
640 x 1536 (5:12 Vertical, the closest to the iPhone resolution)
You can also try other variations, but remember this one thing:

It is important for optimal performance that the resolution is set to 1024×1024 or other resolutions with the same amount of pixels but a different aspect ratio.

[uncut fiber]

can you check size of that model? Just for curiosity? @shy kelp

[pure crystal]

it's not "in the settings"

[uncut fiber]

i mean size in GB

[hardy cipher]

bernix. did you figure out the styler?

[pure crystal]

it's aspect ratio error related

[uncut gull]

I still REAAAAALLLYY want to know how this was made with SDXL guys.

[hardy cipher]

dude, look up deforum

[uncut gull]

This looks like animatediff. But what I mean is I want the original poster to clarify how he did it.

[uncut fiber]

i think i did, unfortunately not portraits and body shots in general. Only silhouette

[uncut fiber]

yes

[hardy cipher]

bernix, you should let me check your workflow. I can fix it up for you

[uncut gull]

Did not see that coming.

[uncut fiber]

i have it exactly as on github. mmnt

is in picture

[hardy cipher]

should work out well

[uncut fiber]

😄 is working

native jugernaut 1536x1536

do you know when you have it opened it performing highres anyway? you must close it (highres)

[uncut fiber]

when opened it perform upscale, to disable it, you must close it. I am in Comfui... Pitty

[uncut fiber]

Seems it will be interesting thing!

[uncut gull]

I just saw a walkthrough of Dracula's castle (historical museum) in VR the day before yesterday. Torture devices. 😐

It looks pretty normal though...

[uncut fiber]

i like it, i like isometric things very

[uncut fiber]

pitty

[hardy cipher]

bernix

if the workflow isn't attached to that for some reason let me know. should be though

just randomized the styles

[uncut fiber]

should be torture room 😄 nice bed for tortured probably 🙂

[hardy cipher]

and used text boxes instead of primitives for the inputs

[uncut gull]

I feel like this would be the easiest style for create usable game assets, but it's trickier than I thought.

[uncut fiber]

o.k.

[hardy cipher]

by all means change it up. just wanted you to see how it works

[uncut fiber]

@hardy cipher i am getting expected results

o.k.

[uncut gull]

The bed looks harmless, but it has a hose inside. It will make it look like you wet the bed. Ultimate torture.

[hardy cipher]

do you have all the nodes?

wasn't sure what you have, but only used one type of custom node I bleieve

other than stylers

[uncut fiber]

i have all

[hardy cipher]

and that milehigh one has over 1500 styles

[uncut gull]

I'm using Isometric gameplay render, blue background, room, *subject* as my prompt, and I'm getting a 100% hit rate for this style.

[hardy cipher]

bernix, ever use ipadapter?

[uncut fiber]

what instead of primitives. Used but not build anything with it

[hardy cipher]

?

[uncut gull]

Even isometric zombie apocalypse. This thing can isometrify anything.

[uncut fiber]

yes

[uncut fiber]

i like pharao

[uncut gull]

Mine doesn't give me realistic renders.

[uncut fiber]

this is styler

just living room

[uncut fiber]

zelda?

[uncut gull]

Shining Force. 🙂 (Probably never heard of it. But it was one of the first games I ever played.)

[uncut fiber]

i am so old, that this is advanced technology actualy. i got my first computer in say 7th class atari 800XL action biker, lost caves and space invanders 🙂

i spend more time in 20th century then this century. But psssst please 😄 to nobody knows it 😄

[uncut gull]

LOL no worries. Congratulations on not being afraid of technology. I'm trying very hard to wrap my head around a future full of AI. We didn't have this stuff when I was a kid. Or last year.

It's 2/3rds full. (...But also 1/3rd empty...)

[uncut fiber]

yes positive aproach!

[west breach]

I think when the inventors for both the internet and email have died of old age, no one is old enough to say they are too old for computers 😄

[uncut gull]

I wonder if the guy who invented spam is still kicking though.

Supposed to be a casino. A little hard to read.

Basically a casino, really... They don't pretend you'll make money, but they do pretend you'll beat the games.

[pure crystal]

@uncut gull try a batman, batcave, batmobile

[uncut gull]

More realistic style than I expected.

[uncut fiber]

he is bit small or car is bus 🙂

[pure crystal]

lol a figurine

[uncut gull]

It's upgraded.

Booty.

[glad grove]

i like that booty

[uncut fiber]

@pure crystal have you comfui?

[foggy garnet]

inside out bread

[pure crystal]

@uncut fiber yes

[uncut fiber]

there is nice styler workflow, you can try

https://github.com/wolfden/ComfyUi_PromptStylers/blob/main/examples/sdxl_prompt_styler.png

[pure crystal]

I'll pluck stuff out of the jsons

I'm a little particular

[uncut fiber]

o.k. this treasure island somehow went wrong

[uncut gull]

LOL that's a landlocked island.

[uncut fiber]

school helper, or how to say

[misty sundial]

Does sdxl work with controlnet as of yet?

[uncut gull]

I need to go to sleep... Maybe just a few more of these...

[uncut fiber]

this is bit better treasure island

[uncut gull]

There's depth, canny, openpose, sketch, maybe some others. Different controlnet models need downloading.

[uncut fiber]

yes just few models released, you need new one

[misty sundial]

more specifally does inpainting work? I hit RuntimeError: mat1 and mat2 shapes cannot be multiplied (616x2048 and 768x320) error. when trying to use inpainting

[uncut fiber]

it means you used model say SDXL for 1.5 thing or other way round

mat1 and mat2 is when not matched models version

[misty sundial]

So is that to say I should use the 1.5 model for inpainting instead of sdxl w/ Controlnet? current running stable diffusion version v1.6.0

[uncut fiber]

@misty sundial if you are working in open outpaint, yes. And try search for some model having inpaint in name. Probably dreamshaper can have. But resolution limitation is not in game when outpainting

it is not about stable diffusion version, it is about controlnet version, which you have i suppose sd 1.5

i am bit lost now, sore 3:30 am almost here

[misty sundial]

is there a controlnet version for sdxl? there is an inpainting module in my current controlnet

[uncut gull]

Age of giants.

Okay. I'm done. Good night all.

[uncut fiber]

i think you have for 1.5 size of models are around 700MB, can you post controlnet models from your folder?

good night!

@uncut gull good night

[misty sundial]

@uncut fiber

[uncut fiber]

oh, you got both in one folder. xl i realy dont know. And yes for inpaint you need 1.5 version of model, best having inpaint in its name. Check civitai.com for it

[ionic gulch]

some of my favorite pics of the last month:
https://imgur.com/a/7Xvu18C

[hoary saddle]

any good workflows for these?

[slender coral]

What's your thoughts?

[hardy cipher]

lol

https://tenor.com/view/dog-doggo-cute-math-formulas-gif-17580986

[pure crystal]

try 1344 x 768, around 30-40 steps and bump the cfg between 6.5-8 for now

[hardy cipher]

hmm

[west breach]

I'd like to see dalle3 follow this prompt: Touareg and Nigerian spaceship pilot with African astronaut helmet with african ethnic patterns, and large senegalese Kaftan and nigerian Agbada, fuji film still from a 1970s indie folk horror cult film by alexandro jodorowsky, dreamy, surreal, outside landscape summer, human face wearing scarlet ibis bird and flower costume made of paper mache and stained glass with a crown, ibis bird mask, Yoshitomo Nara, Gustav Klimt, minimalist

[uncut fiber]

it is still qrmonster for sd controlnet @hoary saddle

[hardy cipher]

that's a lot of things

[west breach]

sdxl is actually really good at following prompts for outfits

[hardy cipher]

interesting. I think I like the results better than the inputs

it's wild that it can fit in all those things. with 1.5 I'd give it beast prompts knowing it wasn't really going to be able to work it all into one image

but made interesting stuff

[west breach]

what the hell is this thing 😄

[slender coral]

The resolution is great, but the cfg is too low for my taste:

[uncut fiber]

some new model of spitfire 🙂

[hardy cipher]

it's amazing how clean it makes things sometimes. really only limited by the resolution

[pure crystal]

7.5 is the sweet spot

[slender coral]

testing with 70 steps, with 7 cfg

[hardy cipher]

triple stacked fighter jet. seems efficient

I only use cfg values between 20 and 60

[glad grove]

michael bay hawx adaptation

[west breach]

Jetfire speeding through the sky during a thunderstorm, splash art render, keyframe art, beautiful, hyper detailed, photomanipulation art

looks like some kind of transformer

[glad grove]

remove keyframe art and photomanipulation art and try again

[hardy cipher]

this is going to make something. not sure what, but something

[glad grove]

he looks like don cheadle 🤠

[hardy cipher]

I have it inserting random styles from a list that has like 1500. always a surprise

and randomized ipadapter effects

ooh

[pure crystal]

holy mountain vibes

[uncut fiber]

so i can download more styles? And how to aply them? @hardy cipher

[west breach]

that's pretty awesome

[hardy cipher]

that setup should be using the mile high one? it has over 1500 in the list

[uncut fiber]

o.k. supposing download json and put it somewhere

[hardy cipher]

well it's already using them I do believe. or it should be

[west breach]

looks like splash art was the term causing that cluster of a jet

[hardy cipher]

I don't even know what splash art actually is

black hole this time. what's that going to do?

[glad grove]

its from games

[west breach]

seems to have a strong effect, without and with 'splash art render'

[hardy cipher]

interesting

I never know what to put in prompts. I just make things up mostly

[west breach]

I use the sandbit file

which are prompts people put in the bot channels

lots of weird shit

[pure crystal]

emphasized the space ship captain part

[hardy cipher]

would that be available anywhere? didn't know that was a thing. assumed they were saved

[prime bison]

Anyone know an upscaler that works better than real esrgan 4x for images with blurry backgrounds? Esrgan is messing it up 😦

[hardy cipher]

got a bit of a bird thing going on here

efreak, maybe ultrasharp? or that other one that I never remember how to type out

[west breach]

#🌠｜show-and-tell message

[prime bison]

Lol

[hardy cipher]

well alright then. sweet. thanks

really rolls off the tongue

[prime bison]

This is my image. (Based from a bot here, upscaled with img2img on Horde using FaeTastic)

[pure crystal]

@west breach great prompt thanks

[prime bison]

Any idea if this is the same as nmkd siax? I can check the actual filename for siax but it's a pain

[hardy cipher]

not sure. I'm really not that familiar with all the different upscalers

[prime bison]

Nmkd_siax.pth probably not then

[hardy cipher]

if the name isn't the same I'd guess no. I feel like they keep adding more to the names as they redo something

[prime bison]

I'll find a Colab to upscale then. Siax is on Horde so I wouldn't have to

[hardy cipher]

is that the money upscaler?

[uncut fiber]

can anybody test studio portrait photo? Getting error, worth to post it?

[west breach]

go on

[uncut fiber]

weird still getting error 😦

[slender coral]

Whats your solution to generating final images in powers of two for compression?

[pure crystal]

I dont have one

[slender coral]

These seem better than the 1024x512...

I guess I can manually scale.

[pure crystal]

so cool

@slender coral give me your prompt

[slender coral]

sci fi, desert shield dome,

[pure crystal]

that's all?

or are you protecting secret sauce

[pure crystal]

they have their own chain technology no one else uses. otherwise, perfect saw

[slender coral]

steps 90, cfg 5.0

[pure crystal]

why does steps 90 cfg 5 feel correct for you?

[slender coral]

Just get more detail I find,

I used to do 20 steps for more stylized.

[west breach]

you can get more detail by using 2 samplers, divide the steps in half, then start the 2nd sampler 1 step later (skipping a step in the middle), but you add that step to the end.

[slender coral]

Can you explain this a little more?

Do you need that node?

[uncut fiber]

is there some easy way to switch same functional nodes easily?

[west breach]

you can use the standard nodes like this

[slender coral]

How much would you say that increases detail? And is that resolution (in the sense of more pixels filling the area)?

[hoary saddle]

do you have a workflow you can share? finished downloading stuff

[west breach]

if you put 1 step at the end, it may return with a sparkly effect for images with smooth areas, but works well for images with lots of detail over the whole picture. if you see the sparkles and you don't want them, just bump up the number of steps by 1 or 2 on the second sampler to get rid of them

[uncut fiber]

it is not for sdxl and not for Comfui.

simply create text, white letters on grey background, type in pos prompt made out of books, Einstein, Library in background.
high denoise strenght and no preprocessor if not needed change threshold or switch white to black or other way round @hoary saddle ping 🙂

i will post one result, hope nobody sees me 🙂

[glad grove]

https://tenor.com/view/i-saw-what-you-just-deleted-gif-22613011

[west breach]

if you want a super easy, compact workflow, try this. you will need my nodes package mikey nodes

it has the skip step and hires fix built into the sampler

[uncut fiber]

to you all 🙂

i have not delete anything 😄

[crisp owl]

I saw it, it was pretty risqué

But, a nice self portrait nonetheless

[uncut fiber]

hyper realism i like it

[crisp owl]

til (after looking it up) that that emoji is a finger and thumb forming a small heart shape.
Never made sense to me before

[uncut fiber]

how much does it cost

[crisp owl]

https://tenor.com/view/thanos-infinity-war-everything-avengers-gif-14066494

[uncut fiber]

it is symbol for how much does it cost

[west breach]

it's popular thing to do in asia

[crisp owl]

Yeah never made sense to me before lookign it up just now haha

[west breach]

you see a lot kpop celebrities doing it to crowds

[crisp owl]

aah

not many deathmetal groups be doing that 😆

[glad grove]

babymetal maybe

[crisp owl]

lol probably

is ladybeard still a thing?

[glad grove]

is that the same as ladybaby ?

[crisp owl]

might have been the groups name. I know the guy was ladybeard when he did wrestling

not sure if the name carried to the group

[uncut fiber]

i will find video one day showing how much does it cost you move index finger up and down...

[crisp owl]

"ladybeard intriduces BABYBEARD" new group looks like, 2021 at least

[uncut fiber]

lets party!

[glad grove]

this looks ai generated

[supple knot]

what are these half rooms called for prompting?

[uncut fiber]

isometric room or anything

[supple knot]

thanks

[uncut fiber]

np

[crisp owl]

hmm, where's my img2img workflow....brb

[supple knot]

lets riff on babymetal pics

[hardy cipher]

I'd assumed it was until I read your comment

[crisp owl]

hurry up dragon pic, I've got an img2img image to work on

[uncut fiber]

oh that dragon 😄

this is mordor tower also in iso @supple knot

[crisp owl]

😆

[glad grove]

🤣

[crisp owl]

oh my

[uncut fiber]

witches

[glad grove]

sexy 🥵

[crisp owl]

I'm dying over here 😆

[hardy cipher]

did reasonably well with the clocks here

[uncut fiber]

clocks working great in sdxl

can you make half past five?

[hardy cipher]

I question whether it'd get that one right

[uncut fiber]

i think it is imposible

[crisp owl]

funky tail, but nice nonetheless

[hardy cipher]

beauty

[crisp owl]

trying to get one breathing out a gemstone/glass fire

blow fire damn you

hmm

[hardy cipher]

close enough

[crisp owl]

guess it kinda works-ish, that's one to gen a hundred overnight and hope for a gold one in the morn

[crisp owl]

glass fire bubbles!

[uncut fiber]

close, seems bit how bubbles from hot soap 😄

[uncut fiber]

how to open fridge? Devs makes mistake somewhere 😄

[crisp owl]

lol from the babybeard image?

[supple knot]

yeah, So many weird ones

[crisp owl]

Interesting you get a lot of fire balls of sorts

[uncut fiber]

german kitchen, get colors under control is nice 🙂

[west breach]

I think I need to train some fire breathing images into my model

[uncut fiber]

i cant make dragon fulscreen...

it was because hyper realism

[hardy cipher]

didn't expect the orange slices here

[uncut fiber]

@hoary saddle does it work?

seems like lemon with orange slice, nice

[hardy cipher]

well they look delightful either way

[uncut fiber]

yes nice

[hardy cipher]

looks like some mint with the candle at the top

those eyes look alright

[indigo vine]

some SEGA Genesis game cover vibes in this one. Like you expect something crazy with lot of actions and you have this:

[uncut fiber]

nice submarines. Whats that game? Is it silent hunter?

oh genesis game then

[indigo vine]

https://en.wikipedia.org/wiki/688_Attack_Sub

[uncut fiber]

all classic names. Amiga, MicroProse

[glad grove]

wish there was a lora with this style

[noble shoal]

If you have a captioned dataset of 20-100 of those i can train one 🤷‍♂️

[indigo vine]

you can try to use the thing with 5-6 images that produces images close to what a lora does

https://www.reddit.com/r/StableDiffusion/comments/16sglm9/no_more_lora_training_only_6_images_needed/

[glad grove]

ill try to make one soon tm

[peak dove]

I need a LoRA which can generate clear sheets of falling water, thru which I can compose characters?!

[noble shoal]

Do you have an example for that?

[woeful patio]

can I use the regulisation images of a real woman when trying to train an anime character lora?

[noble shoal]

I think reg images are not really needed at all. At least i don't use them. But i am sure @rustic garnet has a good answer to your question.

[woeful patio]

it's really unclear since one person says it's vital and another says it doesnt matter lol

[vital ermine]

The answer is yes, but not needed most times. rare cases sometimes

I had a case where all people in my gens were the same. I used reg image of men and gone.

[woeful patio]

Yeah that would be great, really unclear if you can use reg images of people for the training of drawn characters

[hardy cipher]

I think one of the issues might be that two different people train on two different things, or are going for different results. and so what's seen as needed by the first might not be a priority for the second.

and I'd imagine they're less of a necessity when what you're training on isn't super common.

mechbatdragonman

[woeful patio]

cool, Im gonna just train it without the reg images then see what happens.

[vale eagle]

If I have an image and I created depth map based on it, how could I use the depth map to remove the black part of the original image?

without using sampler

[hardy cipher]

I think the sampler would be an inherent aspect of applying a depth map

unless you want to blend it or something

[supple knot]

https://www.youtube.com/watch?v=yE3Chxh7x2Y

[vale eagle]

I want to keep the original image just remove the black part of the depth map

[hardy cipher]

not sure what you mean there

[vale eagle]

For example, keep the character and remove the background

[hardy cipher]

you mean you want to remove the background of the depth map or the original image?

[vale eagle]

using the depth map to remove the background of original image

[hardy cipher]

well if you have access to photoshop or gimp it'd be super easy. and there are background removers as well

[vale eagle]

I want to do it in comfyui

[noble shoal]

Isn't rembg an option here?

I mean without using the depthmap

[hardy cipher]

depth map seems unnecessary in this situation, but what do I know?

[vale eagle]

Yes, rembg seems to work

[uncut fiber]

rmbg working very well

[vale eagle]

After using revision, I noticed that sometime I want the selected part of the image rather than the whole image. I want to remove the unwanted part to get more accurate result.

[hardy cipher]

which one is revision again?

[inland vessel]

i want to train a lora of me but my rtx 2060 can't do it, is there some free way i can do it online?

[hardy cipher]

doubt that

I'd like to be wrong about that. but not sure that sort of thing is offered anywhere

[inland vessel]

:(

[hardy cipher]

might be able to sign up for a trial membership or something somewhere. but training even a lora requires a bit of hardware

[inland vessel]

is there no google collab?

[hardy cipher]

but you'd probably have to pay about 50 cents at most

[uncut fiber]

living room, bedroom, dinning room, bathroom, kitchen, any other rooms? How for example greenhouse that is part of house? We call it winter's garden.

[inland vessel]

i can pay 50 cents but on runpod i have to load up 10 euros

[hardy cipher]

ahh, well there are other options I believe. but not sure what they all are

[vale eagle]

sai revision. Something like this

[hardy cipher]

colab is pretty trash for stable diffusion stuff. it's not banned, but they're trying to do away with guis

ahh, my workflow has a few of those

here's some flowers for everyone

input doesn't really look that much like flowers to me, but I guess I'm not an expert on these things

[uncut fiber]

rose from me

what is that red and white blood cells?

[hardy cipher]

yeah

doing all sorts of things

[uncut fiber]

there isn't better than nature!

[hardy cipher]

I guess it just wants to draw flowers

[noble shoal]

entrance hall, foyer, family room, study, home office, laundry room, nursery

[hardy cipher]

basement, dungeon, wine cellar, spice closet, ball room

[uncut fiber]

library 🙂

Thank you @noble shoal and @hardy cipher

study is where library is. TYVM

wine cellar

what could be this? Garage?

basement?

and according to carpet entrance in white house

[noble shoal]

That's cute. A bit unpractical to live in though.

[uncut fiber]

yes i love it, but dont know what is it. I put all in one so it is randomly picked

could it be atelier?

[noble shoal]

If you have "log prompt" on, you can check what the prompt is.

I mean if you use the styler

[uncut fiber]

i have not. yes stiller, i got more these.

[noble shoal]

With those random styler stuff i write a txt file with the prompt directly into the folder. Do you have WAS Suite installed?

[uncut fiber]

Not sure realy.

i have was node suite

[noble shoal]

Perfect. There is a "Save Text File" node

[uncut fiber]

o.k. will try it!

[noble shoal]

Make sure to begin clean output folder. So the counters line up.

[uncut fiber]

where to connect it sorry i am realy bad in this

text input i did

i have different

[noble shoal]

Can you show?

[uncut fiber]

i have save to file _O

savetexttofile_0

[noble shoal]

WAS Suite -> IO -> Save Text File

[uncut fiber]

o.k. thankyou

[noble shoal]

Make sure those line up

[uncut fiber]

got it and text?o.k going to try it

[noble shoal]

Connect it like here

[uncut fiber]

i dont understand, sorry my ignorance 😄 save image where is output?

[noble shoal]

Send you a DM

[rustic garnet]

in your comfyui directory under "output"

[vale eagle]

a cat theft

[icy brook]

Nice! 🙂

A friend of mine did some similar stuff with my new LoRA - Aether Glitch.

[cursive warren]

Thank you for your kind words.

[upbeat summit]

Hey @cursive warren 🙂 of course!

[cursive warren]

I've been helping out a bit to populate https://rikkar69.github.io/SDXL-artist-study/

[upbeat summit]

oh! one of my favorite resources. nice work!

[cursive warren]

He started out with a list from an existing dataset, I've found quite a few more for him to add, next update will get it well over 5k artists:

https://docs.google.com/spreadsheets/d/1S_4hfLHNx8_AGI_Sa2GR227kUEQUqdmAm7Y3sFXrJZk/

[uncut fiber]

those authors there is list somewhere or is it trial and error?

František Tichý will try those, there is some unknown Tichý in db 🙂

[upbeat summit]

it's research

[uncut fiber]

o.k. then going testing some names i havent search in lists

hey you can add Josef Sudek photographer, tested him and it knows he was b&w photographer

[cursive warren]

This is what I get for that artist:

[wise rampart]

Yamer's Realistic V4 is now on the site:
https://civitai.com/models/127923?modelVersionId=170252

[uncut fiber]

it actualy looks bit like his work, he was about circus, but result i got were not convincing @cursive warren

this is Josef Sudek

way famous then Saudek 🙂

[native knot]

This took a while to build:

[cursive warren]

For many artists, the training data apparently only includes a few random works, often not "typical" ones.

That was my original motivation behind the v2 studies, to document how SDs artist styles shifted between versions, apparently due to different training data .

[uncut fiber]

yes thank you for clarification!

[cursive warren]

You can explore a representative subset of one of the (larger) training data sets here:

http://laion-aesthetic.datasette.io/laion-aesthetic-6pls/

[shy kelp]

with a hint of anime

[uncut fiber]

i think it is beyond my brain cells, but thank you very much

[cursive warren]

For example:

http://laion-aesthetic.datasette.io/laion-aesthetic-6pls/images?_search=Frantisek&_sort_desc=aesthetic

[uncut fiber]

wow!!!

[native knot]

In case anyone wants the workflow. I polished it up just a touch more.

[cursive warren]

No results for Frantisek Tichy, and while that doesn't necessarily mean SD doesn't know that artist at all, it does mean SD probably doesn't know that artist WELL.

[uncut fiber]

https://laion-aesthetic.datasette.io/laion-aesthetic-6pls/images?_search=Sudek&_sort_desc=aesthetic
try art by Sudek and you should get quite similar pictures. Seems he is well trained. 90 rows

yes he is well known probably only localy (frantisek tichy)
Sudek is i believe worldwide

[clever verge]

Long workflow to generate the noisiness? It looks cool.

[cursive warren]

Compare to another somewhat obscure artist, Hugh Ferriss:

http://laion-aesthetic.datasette.io/laion-aesthetic-6pls/images?_search=Hugh+Ferriss&_sort_desc=aesthetic

[dapper dragon]

Used some pictures scanned from paper mostly and prodigy as optimiser, SD XL based model, no noise nodes or post

[uncut fiber]

o.k. @cursive warren

[clever verge]

Good job!

[dapper dragon]

thanks @clever verge

[uncut fiber]

if i can somehow contribute with my free time, let me know @cursive warren would be happy to help 🙂

[cursive warren]

Pretty easy to contribute!

If you are aware of an artist name that works in SDXL but isn't listed yet, add them to the bottom of the spreadsheet.

[uncut fiber]

o.k.