✨｜sdxl

[gloomy lark]

People don't post photos of someone else's underage kid. That's creepy. Just cuz you made it with ai doesn't make it less creepy. It's not artistic, they're not doing anything interesting, it doesn't have a cuteness factor, it's just just looks like someone's kid.

[glad grove]

is that even ai

[fierce heron]

you saying that kids arent cute?

some person you are lol

[gloomy lark]

Reporting.

[tawdry current]

I think we should screengrab this conversation and post it to some place like reddit's AITA or something or some other top 1% channel over there. We can see what a much larger echo-chamber thinks of this conversation.

[glad grove]

well a redditor like you would def do that

[tawdry current]

I don't have an account on reddit and never have, but I'm aware of how things are over there

[fierce heron]

please

why the fuck would i care lol

[glad grove]

better watch out he gonna call the reddit bureau of investigation

[gloomy lark]

@smoky patrol

[glad grove]

yea better call the mods and ask them where is the guy who posted the sex pics

[fierce heron]

im sorry. kids are not cute. they ugly

im making ugly kids

[minor pecan]

[glad grove]

better ban that emoji as well

[fierce heron]

@tawdry current for u

[glad grove]

she looking fine and rusty

[hoary vapor]

pedo chat turned into gilf chat?

[tawdry current]

And those are fine... It's just highly odd for a "64 year old man" to be sitting around all day making pictures of what he calls "cute kids." There's no version of reality where that isn't frowned upon and viewed as a sign of things...

[fierce heron]

im a woman so its okay now?

my name is Julia

[glad grove]

why are you telling us your life story here

[tawdry current]

No, no adult should be sitting around making pictures of little kids.

[smoky patrol]

Guys, please move on from this conversation. Thanks!

[fierce heron]

i wasnt trying. i was making 18 year olds with "kid, child, preteen" in negative prompt

[glad grove]

if you see any little kid pic like that @tawdry current claims you can tell us please so we can understand his logic

[fierce heron]

to remind me of what it feels to be young 👵

[tawdry current]

So you're just going to allow people to post hyper-realistic children on here? Did the parents of the children used in the images that trained the models consent to allowing their likenesses to be used by people around the world to make AI children? Since you're a mod and all, we should know the exact stance on this topic.

[glad grove]

go outside redditor

[tawdry current]

I guess I can just report this discord channel to discord itself then and let them sort it out

[glad grove]

oooh

https://tenor.com/view/hhh-hbk-wwe-cringe-dx-gif-10349784

[smoky patrol]

Like I said, please move on from the conversation. I volunteer my time here; I love AI. If you have questions about the rules, you can view them via #✍🏼｜rules-and-tos

[glad grove]

lets hope he doesnt call the manager

https://tenor.com/view/customer-service-karen-skboz-st-louis-i-want-to-talk-to-the-manager-gif-17655801

[fierce heron]

crazy how much using the fooocus ksampler has improved my picture quality

[glad grove]

wth are those amish eatin

[native knot]

Damnit...you beat me to a manchild post by moments.

[fierce heron]

average redditor

[copper kraken]

does this mean that only extreme gore and blood (guts, disgusting blatantly R-rated stuff) is forbidden, and that moderate stuff that could maybe just barely earn a PG-13 rating is okay, like some smudges of blood on a face, or some open cuts on a soldier fighting his way through a jungle or something?

[fierce heron]

i think "excessive" means PG-13 blood would be okay

[copper kraken]

this is a question for @smoky patrol as that is pretty ambiguous... i guess my real question is... does this mean zero blood/gore? as in G-rated?

i heard it meant none

[fierce heron]

imaging banning nose blood 😆

[glad grove]

just do it like some tv shows in china and make all the blood white

[fierce heron]

sd3 could probably do it

[gloomy lark]

yuck

@smoky patrol may not have done anything, but this content directly violates Discord's own rules, so it's now reported to them instead.

https://twitter.com/StabilityAI/status/1783518974090858602

Stability AI (@StabilityAI) on X

Our SVP of Integrity, @ellagirwin, will be joining @Thorn, @AllTechIsHuman, and other industry leaders for a livestream on April 25 at 12PM PT/3PM ET. They’ll discuss the Safety by Design initiative announced this week and other proactive ways to mitigate the spread of harmful…

Twitter · Apr 25, 2024, 3:30 PM

You'd think Stability would take a stronger stance on this.

[glad grove]

prob cuz no one is posting abuse here

[gloomy lark]

suggestive sexual content of minors violates Discord's rules. What he posted directly does it.

posting underage kids in bikinis is gross.

[glad grove]

well if i could guess age of pictures id make lots of money

[gloomy lark]

this reminds me of the discussion i had with a model maker on tensor art, where they said that pedos should have an outlet for the stuff they want to look at. kept deflecting.

[glad grove]

keep projecting

[gloomy lark]

anyone who's had kids knows how old those kids are.

[glad grove]

its ok you can let it all out here,all the heavy stuff that been bothering you for years

[gloomy lark]

keep deflecting.

[glad grove]

keep projecting

[gloomy lark]

fine, I'll post his message on twitter, lets' bring some real attention to my projecting.

[glad grove]

you can report him to twitter if you want,thats on you

[gloomy lark]

not reporting him on twitter, posting a link to his message and the mod response to it.

[glad grove]

what mods

[gloomy lark]

@glad grove I think you're right actually. I think this person is a fed. they're a glowy. it's all just a honeypot trying to figure out who will upvote their gross underage crap and so they can out them. I was wrong to get in the way.

[glad grove]

yes they are a honeypot,they post non sex pictures to trick you into wanting the bad stuff and then they can put a chip on you so you swear loyalty to lizard alien hillary clinton and can now control clouds with haarp

[gloomy lark]

that you're saying what they posted isn't sexual, is kinda messed up when I'm looking directly at it. you know they posted undersage kids in bikinis and school girl outfits, which is directly mentions in the rules and tos right?

that's some serious gaslighting right there

[gloomy lark]

This includes NSFW content involving or depicting underage characters such as schoolgirls, lolicon, etc.

[glad grove]

indeed the nsfw thing

[gloomy lark]

i know i'd get fired if I brought that up at work

[fierce heron]

sorry i missed the no schoolgirl thing in the rules. ive removed the image

[glad grove]

dw i heard u can get executed in iran for wearing this

[dim knot]

my try

[dim knot]

blue didnt work

[glass forge]

I think we inadvertently created a synthesis of the ultimate beauty ideal. by training so many faces we have the most generic pretty asian and most generic pretty American girl archetypes.

she is everybody and nobody at the same time

[dim knot]

literally me fr

[glass forge]

you made a lora of yourself

but if you just type in girl ull just get those synthesized archetype faces

[dim knot]

no i have the same uniform,only difference between me and her is her face,the hair,the skin and 200 kg

[glass forge]

damn

[tawdry current]

This is what I'm going to do from now on when people flood these channels with boring waifu non-sense

[dim knot]

sorry i flooded the channel that belongs to you

[glass forge]

the ponies are coming!

[fierce heron]

keep em coming

[glass forge]

what checkpoint is this?

[dim knot]

is that juggernaut with real lora

[glass forge]

not the borign ppl lora?

😄

[tawdry current]

out of the nearly infinite number of things that can be created with SD and people just sit around all day making waifus lol... tragic...

[fierce heron]

reality engine + boring reality_facesV4

[dim knot]

even more tragic that you care about it

[fierce heron]

is there another lora thats good at this?

[dim knot]

i dunno i use that one only

[fierce heron]

boring reality?

[dim knot]

yes

[fierce heron]

cool. it's really good

[glass forge]

it can get scary

too real.... if you know what i mean

[fierce heron]

thats what i like

[glass forge]

dont promt like indian man in public bathroom and stuff like that

now why would anyone try to promt for that anyway right?

or cemeteries

sweat shops

[dim knot]

skibidi toilet india funny gyatt

[glass forge]

boring lora has been to places....

[fierce heron]

im gonna take that as a do it

[glass forge]

good luck

[primal pine]

guys can this server help me to generate some simple lined drawing pica?

[dim knot]

theres a skibidi lora in case you want to try

[primal pine]

how?

[gloomy lark]

sorry, they're too busy talking about underage girls and shit porn.

[primal pine]

ahahahahahahah thats normal let em have fun

[dim knot]

needs a school uniform

[primal pine]

so guys any suggestion??///

[copper kraken]

hold the line, hold the line! reinforcements are coming

almost done training a lora here

[gloomy lark]

to think I was troubleshooting someone's python code on this channel earlier.

[fierce heron]

can you give an example?

[glass forge]

Isn;t there anything Steven Seagull can;t do?

[copper kraken]

RX808 and i had a great conversation about unsampling etc

i learned some really interesting shit as a result

it was great

[gloomy lark]

how do we rewind this chat log about 5 steps?

[dim knot]

was that before he went karen mode or after

[primal pine]

so?

[dim knot]

you need a lora for that

[tawdry current]

i'm in my mid 30s, i can't be a karen yet.

[dim knot]

you can always identify as one

[primal pine]

i have a nvidia geforce 920m i cant run ds or comfyui

[dim knot]

ok then im sorry

[primal pine]

so i have no option?

[dim knot]

🤷‍♂️

[fierce heron]

you can buy credits on civitai if you dont have a PC to gen with. there you can use various LORA to make stuff like this

or maybe even SD3 can do it

[primal pine]

how much it cost? whe it stand for money theres a lot of option tbh

[fierce heron]

havent looked into it.

[primal pine]

okaaay guys thank u so much

have fun

[fierce heron]

you guys say my work isnt art so imma use an art lora to make it art

[dim knot]

better add a trigger warning,for the karen

[fierce heron]

oh god... its just making it more nsfw

[dim knot]

here it only gave her a tie

[minor pecan]

that is HOT

[fierce heron]

---> #🆕｜sd3

ur cheating

[fierce heron]

in here we make art only

[dim knot]

now make Zendaya eating those greens

[glass forge]

i bet there is a steven seagal lora

[dim knot]

there is a few,i also seen hillary and one from the queen of england

[fierce heron]

theres probably a whole checkpoint

[grizzled verge]

But you have to be a true ai artist to get access

[fierce heron]

epic tennis racket win

[dim knot]

selena williams if she was good

[glass forge]

poor guy, he wishes hw could have worn such nice outfits 50 years ago

[glass forge]

I can tape a banana to a wall...

I am a true (ai) artist

just like real artists had to go through the hurdle of proving they are artists even tho theyre not just painting old ass aristocrat assholes now AI artists will have to prove that just coz they use AI they can still be artists

hopefully this time wont take nearly a century to do so

[tawdry current]

AI makes art, sure, but it's not art made by a person. Prompt and workflow engineering is a thing, being an AI artist is not.

[glass forge]

its inevitable....

it will mature into art

[grizzled verge]

Watching music artists react to how ai music is evolving like images did is hilarious. The copium is real

[glass forge]

yes

photography was blasphemous too coz it rendered a lot of painters irrelevant unless they moved on to soemthing else othern tan painting old ass aristocrat assholes

[tawdry current]

I do a lot of traditional, digital, sculpting and 3d sculpting artwork. I can say with 100% certainty that the things I make with AI are not my own art. In the world of commission artwork, if I have someone paint me a painting, where all I did was tell them what to make and maybe give some feedback throughout the process, I don't get to claim I made the painting. Same goes for AI art.

[glass forge]

a skill that took alifetime to master for centuries and centuries now was replicable even better with one click

fk u photographers what is this new sorcery!

[dim knot]

furniture too,ppl used to do it by hand

[copper kraken]

a film director is still an artist... it's just a diff role

[glass forge]

so when u use a collage and bash togetgher some art using clearly other peoples art

isnt that ur art then?

[tawdry current]

correct, but they don't get to take home best actor of the year awards

[dim knot]

they take best director instead

[tawdry current]

which means they are up against other directors, not actors

[dim knot]

some actors also direct

[tawdry current]

still two separate categories for awards

point is, AI is the artist, you are not. there is no beating around the bush on this

[dim knot]

thats your opinion not a fact

[tawdry current]

you're just paying commission in the form of a power bill

[glass forge]

ai is just a tool

[grizzled verge]

By that metric we just train ai on ai art and the ai artists will become artists

[dim knot]

well so is he but what about it

[tawdry current]

i'm a big fan of the band

what can i say

at any rate, people can say whatever they want about AI art and content creation, in the little echo-chambers, doesn't change the opinion of the vast majority of people on the planet. leave your house for a day and randomly poll people in a mall or grocery store. explain to them the very basics of "ai model do thing cuz trained on stolen data of trillion image," then you type a prompt about "maek waifu plz" and then an image comes out. Ask them if they believe that makes the human operator an artist or not.

[glass forge]

then train your own lora on 100% copyright free materials and do somethign totally unique and awesome

[tawdry current]

copyrightless or not, you can explain that part too...

[dim knot]

he is the guy who said he could sue you for making your own loras

[tawdry current]

the answer is going to be the exact same: no, it doesn't make that person an artist

[dim knot]

since he speaks a lot about echo chambers he seems used to being in one

[glass forge]

so if you use photoshop are you an artist

or you have to use totally non digital means to be called a real artist?

[rotund surge]

Who said writing prompts makes you an artist?

[grizzled verge]

Yeah and all those people are gonna consume lots of ai art and music giving zero fucks about it because they enjoy it. It's all talk

[glass forge]

where do we draw the line

[rotund surge]

Also who said that this models produce art? They just produce images

[glass forge]

only real physical materials make you an artists huh

[copper kraken]

there is no such thing as art

[glass forge]

lol

[dim knot]

there is,there was a banana on wall thats art

[glass forge]

:))

[grizzled verge]

It's virtue signaling. The ai artists will end up winning. When noone can tell the difference and they can do it for a fraction of the price

[rotund surge]

A banana sticked in the wall is art for some people in fancy museums so....

[dim knot]

also people who disagree with me = echo chamber

[tawdry current]

digital artwork is still artwork because you had to perform all the tasks. layers, blending, yada yada yada. still had to touch pen to tablet to draw stuff or click the mouse. that's why someone who does digital artwork is called a digital artist. ai art is digital art, but you aren't the artist, it is. you just tell it what to make for you and wrangle some stages. no different than giving feedback to a commission artist.

[dim knot]

the climate protesters throwing poop at famous old painting also counts as art because they are using their hands

[tawdry current]

and that shit sells

havent you ever seen turd installations

[dim knot]

sorry havent read kotaku yet

[tawdry current]

lol

[glass forge]

idk man anyone who tried to produce anything good other than pump out waifus will tell you it takes time and effort to get good stuff out of ai

but whatever

i guess your barometer is effort to be an artist

[copper kraken]

now what if all you do with AI is a 0.2 denoise to clean up artifacts from your traditional illustrator/photoshop workflow

are you an artist?

[glass forge]

Well in ai you had to make promts, refine the shit out of them inpaint outpaint probably bring it into PS to final edit so ...

[grizzled verge]

It's just different tools that does things for you so you don't have to do all the work.

[tawdry current]

trial and error, sure, but they still didn't make any of it: a random seed did. i make a ton of stuff with SD but i'm not going to lie to myself and say it's my art. idk, maybe because am an actual artist as well, there's no way i could lie to myself like that.

[tawdry current]

no, it does all the actual work

[glass forge]

no matter what tools come along it will always take skill to produce anythign good

[tawdry current]

you don't take credit for a commission artists work, just because you sat with them for hours guiding them on what you want. even if it was through every stage of the painting or sculpture

a backseat driver isn't the driver

[glass forge]

well y'know everythign is first laughed at then resisted than accepted than becomes the norm

guess where we are with AI

:DDD

[copper kraken]

what if all you do is have them lightly airbrush your painting

[grizzled verge]

Nu-ugh I made the prompt. I clicked the generate button. MY WORK

[glass forge]

exactly

you have to at least fix the hands! :))

[grizzled verge]

There are conscious human decision making at play

That it doesn't do for you

[tawdry current]

see what i just said about commission art

[glass forge]

fixing hands alone is a chore

sigh

[copper kraken]

is a screenwriter an artist

they don't act the scenes, direct the shots, edit the film, write the score, or anything

[dim knot]

no because he only writes stuff on the screen

[copper kraken]

and half the time, the screenplay is heavily edited on the fly anyway

[gloomy lark]

so we are writers.

[glass forge]

and the award for best promt goes to...

[gloomy lark]

llama3

[copper kraken]

and there's a "slippery slope" too with image editing, which is what i love doing

[dim knot]

the kotaku writer

[glass forge]

[copper kraken]

my goal is to get to the point where i have total control over the composition and i'm not really using prompts for a whole lot beyond quality, tone, etc

[dim knot]

just dont post any female image or you will trigger it again

[copper kraken]

to be able to use it like a photoshop on steroids

[grizzled verge]

When a whole ass movie is based on a prompt this might as well become the case

[gloomy lark]

there you go conflating again.

@copper kraken sd3 is gonna be great when we can get it local. prompt: some completely weird stuff that makes no sense but is really cute

[dim knot]

https://tenor.com/view/tyler-rhoades-andrew-ttrpg-rpg-dnd-gif-12471387

[copper kraken]

no kidding

so here's another thought experiment:

"generate a cool cyberpunk city" does'nt make you an artist...

[fierce heron]

guys all chill out and look at this cutie

[copper kraken]

but what if your prompt described literally every last detail, down to the angles, dimensions, individual pixels, hues, direction of lighting, hairstyles, weight, angle of the face, crookedness of noses, i mean everything

then are you the artist?

[gloomy lark]

what even is this.

[dim knot]

https://tenor.com/view/no-nooooo-sad-cry-crying-gif-7599369

[copper kraken]

thats so awesome haha

[dim knot]

theres also a stellar blade lora if you wanna try it out

[fierce heron]

looks like there's 10

[dim knot]

oh they are fast

[fierce heron]

probably better just play the game if you want some hot spandex action

[dim knot]

i want to but im scared the karen may cancel me on the twitter

[dim knot]

https://tenor.com/view/yes-sir-gif-13152636

[fierce heron]

game itself already got cancelled. they also censored it

probably to get a PG rating or whatever

[fierce heron]

hey at least we not paying for em

[gloomy lark]

I get my money's worth.

[lavish parrot]

wow

[fierce heron]

i tried

[minor pecan]

Nice peace

[glass forge]

anti-waifu pics

[glass forge]

🤣

[stone fossil]

😛

[fair idol]

image of jasmine flowers

jasmine

[copper kraken]

Here is the image you requested.

[uncut steeple]

PixArt+SDXL refiner

[south horizon]

wow these are quite interesting... combining multiple outputs?

[south horizon]

you got some good glasses on a few of these -- did you prompt anything other than 'glasses'?

[stone fossil]

No just text 2 img.

[fierce heron]

Nope, but I did facedetailer at 0.40 (karras)

[south horizon]

Ah cool, different details plus text, pretty good results

[south horizon]

Cool, I always get ugly glasses 😝

[gloomy lark]

I like to badger people

[native knot]

I just want to mention that the Wu-Tang Flan ain't nothin' to fuck with.

[glad grove]

that flan slaps 🍮

[copper kraken]

holy shit all of paincreator's models on civitai are gone

[copper kraken]

no more darkarts

[fierce ruin]

this is impeccable

[near jackal]

Ten-year-old boy, deftly breaking into the intercom in the entranceway

((outstanding quality)), ((masterpiece)), ((realistic)), a boy deftly breaking into the intercom in the entranceway, his figure blends in perfectly with his surroundings, creating a picture that captures the viewer's attention, a pleasing picture, a masterpiece.

[humble cape]

@smoky patrol a trap link by @tacit ferry

[smoky patrol]

it's a trap!!! Runs

[humble cape]

[rich olive]

Photorealism from ideogram

[uncut gull]

That poor centaur lost his hand.

[rich olive]

I didn't even notice that lol

[rich olive]

Base sdxl no refiner no hirefix or upscaling

[noble shoal]

Base is an awesome underrated model

[gloomy lark]

try cosxl as a base. it's definitely a newer model with different training and of course the higher color range. I've had bad luck merging it with other models, it changes too much of their feel no matter what I try in the merge block options, but as a standalone model it's worth looking at.

[tawdry current]

Are you making sure to the right sampling for it? In comfyui, you gotta use that continuous node between the model and sampler. I think 200 is the max before it starts to fry things. And usually, you need a cfg that's a lot lower. Like 3-5

this node

for cosxl, you should have the sigma_max raised to like 150-200

[gloomy lark]

? i do? for the models I've merged together, I'm not using that and the output looks great. it's just that it's not as prompt following as my original model before the merge.

I'll go back to the original merge workflow I downloaded and see what's up with that and see if it changes the output at all.

[tawdry current]

yeah it probably also depends on things like the merge ratios as well, when you merged the models together

[gloomy lark]

right. definitely worth a try

working on that new pixart smaller t5 model first.

it's going to speed things up a ton

[tawdry current]

Yeah I'd give it a shot. I think the sampler would just see your cosxl merge as a normal sdxl model and would like do the typical discrete 14ish max sigma for it that gets used for regular sdxl models. I haven't played with cosxl merges much, but I'd assume you'd still want to do continuous sampling with a higher max sigma. Like when you use playground2.5, the samplers know and automatically do the continuous sampling with like a max sigma of 80 by default. I don't know if they can automatically detect cosxl merges like that or not

So it would make sense that your merges don't give the greatest results if the sigmas and sampling aren't being done correctly

(oh and pg2.5 isn't a cosxl merge, but they are very similar in what they do)

[gloomy lark]

yeah, so in the merge workflow published by comfy, he's got that continuous edm node in there. even in there, I can see that the output is way less prompt following than my original model. so although I should have those node running in any cosxl model workflow, I think it's just the cosxl being way less prompt following in general than the ones I'm getting off civitai which have obviously gone through endless finetuning compared to what stability has as base.

I've even tried putting output block 0-4 back to 1, and it's still not as good as not merging it at all. it's really a bummer because I want the extended color range. it looks massively better with that.

[white vigil]

a futuristic setting where a woman with blue hair is lying down on a mechanical bed or platform. The woman has several mechanical and robotic attachments on her body, including a chest plate, robotic arms, and a headpiece. The room is filled with various technological equipment, wires, and machinery. The overall ambiance suggests a high-tech environment, possibly a medical or research facility

[glass forge]

What does "clip skip" do actually?

[native knot]

Read this:
https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/5674

[glass forge]

ok ill leave it at 12 then

:))

[scarlet sequoia]

Close up of a lotus flower, outline, black and white, hand drawn

[stray warren]

Smell the poppies

[cyan crown]

https://civitai.com/models/437363?modelVersionId=487195

[icy brook]

Just released this light painting LoRA.

https://civitai.com/models/410151/aether-light-lora-for-sdxl

[sterile hazel]

https://tenor.com/0Yin.gif

[cerulean dove]

Hi Guyz i am new here can any body help me to prevent text generation on the image from stable diuffusion

tell me if can we finetune the Sdxl for not producing the text on any image, becasue stable diffusion is very bad in producing text so can we somehow stop sdxl to product any kind of text on the generated image, Currently i am using negative prompt

NEGATIVE PROMT = "text, fonts,words,3d, cartoon, anime, (deformed eyes, nose, ears, nose), bad quality,bad anatomy, ugly"

but it does not listen to the negatrive prompt so well
i want to generate canva template for christmas halloween etc with no text on it but it always put text with wrong spelling

[meager canopy]

Why do all of your example images say they're made with DreamShaper XL?

[native knot]

Because that was the model he used?

[meager canopy]

He's using them as examples from his own model though...which has a different name.

[native knot]

He's posted a LoRA, not a checkpoint.

[meager canopy]

Doh! I should've looked closer 🤣🤣

Sorry @cyan crown

[cyan crown]

Lora yes.

Some also with Realvis but it depends

[fierce heron]

(im awaiting 80yo men in schoolgirl uniform)

[minor pecan]

Can you make strong business women

[fierce heron]

whos watching she hulk?

[tender storm]

A humble man standing holding a recycling box and smiling inside a factory.

[minor pecan]

[wet nacelle]

https://tenor.com/view/cat-kitty-cat-staring-i-am-you-im-you-gif-26272604

https://tenor.com/view/jerma-jerma-yell-jerma-yell-orange-jerma385-twitch-gif-26524964

[wicked portal]

What are you using for this? It looks great

[fierce heron]

realism engine and boring reality faces+DPO lora. models for lora at 0, clip at 1

[wicked portal]

Thanks man!

@fierce heron What do you mean by "models for lora at 0"?

[wicked portal]

Oooh I see, thanks for your help

[fierce heron]

np 🙂

[wicked portal]

How do you make your images? The details are so clean

[glass forge]

Not sure if high tech or slum.

[wet nacelle]

https://tenor.com/view/buffoon-gif-21598206

[wet nacelle]

Download my image and open it in a text editor. I keep my metadata on my images.

Ask me questions about anything you have issues with.

[wicked portal]

@wet nacelle Fantastic, thank you

[wet nacelle]

yuppers

[wicked portal]

@fierce heron Got your output style working too

[wet nacelle]

nice

[wicked portal]

Wow that crystal clear model works great even without an upscaler, thanks for the recommend

[glad grove]

we need more uniforms

[noble shoal]

Uniforms? Sure thing.

[meager canopy]

https://tenor.com/view/wtf-gif-22886423

[tawdry current]

omg don't give me any more ideas for trolling the thirst spammers that only post waifus...

[noble shoal]

A Blobfish is a nice creature for that. 😏

Especially those with blonde wigs

[tawdry current]

i'll keep that in mind lol

[tawdry current]

Reminds me of doc Manhattan, maybe throw that into the prompt for more glows and to make it look even more ethereal

(super hero from watchmen)

[noble shoal]

[meager canopy]

https://tenor.com/view/animated-thumbs-up-gif-25581782

[earnest hull]

Vermeer style girl with pearl earring, cute minimalist cartoon style, bust, 12 different poses and emoticon combinations, emoticon pack size as big as, Vermeer style sticker set emoticon pack with various emoticons, (angry, happy, angry, pouting, thank you, hahaha, blessing, good night, love, flying kisses), etc., as a set of illustrations, bold lines, cartoon style, cute style, sticks character style, dynamic pose, white background, f/64 set, related characters, Family guy style

[tawdry current]

I don't know if these are waifu enough or not for the people here. I tried to throw in stuff like "japanese university uniform" into the mix

[noble shoal]

Pretty waifu already I would say. I think there are a few people that would like that.

[tawdry current]

I would say whip out the flood it again meme with McConaughey smoking a cigarette photoshopped as Jesus, but these reverse mermaids would survive just fine

[gleaming mango]

https://tenor.com/bAFga.gif

[gaunt tapir]

mind sharing the prompt? I only got this far 😂

[tawdry current]

I'll have to dig it up, I used pixart and then resampled with sdxl at like 0.6 or something to mostly keep the overall forms

[tawdry current]

And I should break my moral stance on face swapping people to do a high quality version of the meme. I find myself reaching for it often in this channel

[ashen otter]

Please do so

[wet nacelle]

WE NEED MORE GIRLS!!!

[wet nacelle]

no probs

I couldn't believe it either.

[mighty beacon]

guys help me out

[wet nacelle]

No

[mighty beacon]

[wet nacelle]

https://discord.com/channels/1002292111942635562/1002602742667280404

[tawdry current]

The prompt for the reverse mermaids was "a full-body portrait of a woman with the head, face and neck of a blobfish, human-fish hybrid, wearing a japanese university uniform, subtle gradient background, soft focus, warm glow" and I remember it taking around a dozen seeds to find a decent one. It kept doing what yours did where it would just put a whole fish where the head was.

but like i said, it was a pixart to sdxl workflow. was mostly just using pixart for the better prompt cohesion and to lay the scene out

oh and no negative prompt. i rarely ever use them

[twilit grove]

ai getting too real

[elder plume]

i get this "half body" bug a lot

[fierce heron]

just what you all need

[hoary vapor]

Belly button piercing combined with zipper? Hm, thats kinda practical

[mossy latch]

How would I even begin to make something like this lmfao I love the quality

[frail scaffold]

hello can someone help me with a problem?

i download two models,and there are all 1.0 model

but when i use kohaku model, the whole workflow will notice error

the other one is good to use,why?

[native knot]

You need to share what the error is.

[frail scaffold]

oh

ok can you have a look?

[native knot]

It's literally telling you the problem at the beginning of the error. You're trying to use an animated diff model that is for use with SD1.5 models. The two above that you provided are a LoRA and an SDXL checkpoint; neither are compatible.

You can see the model type in the 2nd screenshot you posted:

SD 1.5 models will show SD 1.5 in that place on Civit.

The other one is a LoRA, not a full checkpoint. LoRAs are applied on top of models.

[frail scaffold]

i konw it is version problem,but just dont feel it right~~

[native knot]

Feel has nothing to do with it. Your error tells the issue. You cannot mix 1.5 and XL models.

[frail scaffold]

this work i combined these 2loras, they have different version, but can work very good,

but i cant change to kohaku checkpoint??

what can i do if i really wanna try kohaku checkpoint?

[gloomy lark]

there aren't a thousand anime checkpoints for 1.5?

[native knot]

Gongbi + Chinese ink comic-strip = ok (both are SD1.5)
Ink Wash painting + Kohaku XL = ok (both are SDXL)

But your error above talked about animateddiffMotion_v15.ckpt. That is only compatible with SD1.5 models.

[worthy crane]

hello

[frail scaffold]

ohhhh i make a mistake, thanks

i konw what is wrong with the model~~thanks for you help@native knot👍

sorry to say that but the problem still remain~~xi_shuimo is download from Ink Wash painting, it didnt work~@native knot

[native knot]

Again, your error is telling you that you cannot use "animatediffusionMotion_v15.ckpt" with the SDXL models. Somewhere in your workflow you are still trying to use that.

[frail scaffold]

here?

this model cause the problem?

[native knot]

Yes, the one you had selected was v15. You should be trying an XL version.

[frail scaffold]

cause another problem~~

than what is this?

i alredy thange to sdxl

[native knot]

You need to look through the rest of your settings for SD1.5 stuff.

[frail scaffold]

i can load my workflow,can you help me have a look?

[native knot]

Are you trying to use controlnet somewhere in the workflow?

[frail scaffold]

the whole workflow

yes depth

[native knot]

Yup. Controlnet model for 1.5

[frail scaffold]

what?

[native knot]

You cannot mix 1.5 and XL. Go through and change them to SDXL compatible controlnet settings.

SD1.5 and SDXL do not mix. Period.

[frail scaffold]

hahaha

thank

[native knot]

Yup

[frail scaffold]

finally right thanks

very helpful

👍

[lofty juniper]

how to make image

[native knot]

This discord does not have a bot. So you either go to a website somewhere or download the software to do it on your computer.

[copper kraken]

Here is the image you requested.

[gusty elk]

在一个宁静的小镇上，住着一个名叫小杰的少年。他拥有一双充满好奇和探索的眼睛，总是渴望着冒险和发现未知的世界。一天，小杰偶然发现了一本破旧的地图，上面标记着一个神秘的岛屿——传说中的梦幻岛。
梦幻岛被描绘成一个充满奇幻生物和美丽景色的地方，据说只有最勇敢和善良的人才能找到它。小杰的心中充满了渴望，他决定踏上寻找梦幻岛的旅程。
他背上背包，告别了家人和朋友，踏上了一条充满未知和危险的道路。在他的旅途中，他遇到了各种各样的挑战和障碍，但他从未放弃。他学会了与大自然和谐相处，帮助他人，并结识了一些忠实的朋友。
终于，经过无数的艰辛和努力，小杰终于找到了梦幻岛的入口。当他踏进岛上的时候，他被眼前的景象所震撼。五彩斑斓的花朵，高耸入云的树木，还有那些他从未见过的奇幻生物。这里是一个真正的仙境。
小杰在岛上探索了数日，每天都有新的发现和奇遇。他与岛上的居民们结下了深厚的友谊，并学会了更多关于勇气、友谊和善良的道理。
然而，他渐渐发现，梦幻岛正面临着危险。一股黑暗势力企图摧毁这个美丽的地方，并剥夺岛上的生物们的自由和快乐。小杰决定不再袖手旁观，他带领着他的朋友们，与黑暗势力展开了一场激烈的战斗。
经过一番惊心动魄的战斗，小杰和他的朋友们最终战胜了黑暗势力，拯救了梦幻岛。岛上的生物们重获自由，他们用欢呼和感激的目光看着小杰。
小杰明白，他的使命已经完成。他带着满满的回忆和成长，返回了家乡。他知道，这段旅程不仅仅是一次冒险，更是一次心灵的洗礼和成长。
从此以后，小杰变得更加勇敢和善良。他用自己的经历和智慧，帮助他人，传播爱与和平。他的故事也激励着更多的人，去追寻自己的梦想，勇敢面对困难，成为一个真正的英雄。
这个故事告诉我们，勇气和善良是我们最宝贵的财富。只要我们心怀梦想，勇往直前，就能创造出属于自己的传奇。无论是在现实世界还是在梦幻岛，每个人都有机会成为改变世界的英雄。

[woeful patio]

what is the best method of captioning images right now?

[tawdry current]

i've tried them all pretty much and found llava 1.6 mistral 7b to be the most reliable. moondream2 is pretty reliable as well. there are some newer ones out there as well that try to mix the vision part with other llms, like i think there's a llava llama3 model as well, but i think it uses the older llava 1.5 vision model

they all have their pros and cons though, i just found mistral variants to behave themselves with decent system prompts

[copper kraken]

Here is the image you requested.

[regal folio]

我试试看能不能收到图片

生成一个招财猫相关的app图标,尺寸为1024*1024,卡通元素多一些,把整张图片都填充满,不要有空白

Generate an app icon related to Lucky Cat, with a size of 1024*1024 and more cartoon elements. Fill the entire image without leaving any blank spaces.

希望能收到图片

[frank kraken]

how do i start usingbthis AI????????

[native knot]

You download, install, and configure something like ComfyUI, Automatic1111, or Fooocus. Then you download model files and put them in the correct folders. Then you make choices in the program's UI by clicking buttons and typing words.

Do all that successfully and in the right order, add a sprinkle of patience, and you get images.

[meager canopy]

You need a decent GPU as well

[native knot]

Not necessarily, but it helps. You can generate on shitty GPUs and CPUs, but it's super painful and everyone that tries it hates themselves.

[meager canopy]

...and the rest of the world, especially those with high-end hardware 🤣

[native knot]

Truth.

[frank kraken]

i have a 4050 is it good???????

[meager canopy]

Should be good enough 🙂

[frank kraken]

i checked a tutorial i tells me to istall python and git

[meager canopy]

Depends what you want to do

If you use ComfyUI, then it's a single install which has that built in

https://github.com/comfyanonymous/ComfyUI

[frank kraken]

thanks men ill check it out'

[meager canopy]

Install instructions https://github.com/comfyanonymous/ComfyUI?tab=readme-ov-file#installing

[frank kraken]

❤️

[native knot]

Egg:

[meager canopy]

https://tenor.com/view/correct-gif-12268136301219154789

[gaunt tapir]

awsm, thanks!

[native knot]

Egg?

[stray warren]

Cheers!

[minor pecan]

Replace b with C!

[humble viper]

https://tenor.com/bQ3gZ.gif

[noble shoal]

CIG?

[minor pecan]

[molten gull]

what sampler do you have to use with 1.5 models ?

[pallid lynx]

pretty new to this. what are these nodes called and what do they do?

[native knot]

Those are called Reroute nodes. They're simply for helping to guide your noodlez around. They're not required if you don't like them, but some do.

[pallid lynx]

ah ok. they are just for organization?

[native knot]

Yeah. The most efficient use is something less obvious, though.

Let's say I have 2 nodes that are going to receive the same input and I'm going to group them...add a re-route node and connect the 2 other nodes to it instead of 2 noodles directly to the 2 nodes. Then include the reroute in the group so you only have one input for the grouped node.

[meager canopy]

Why does the Pixart-Sigma 512 model come out like this?

[copper kraken]

are you using the wrong resolution

[meager canopy]

Yes! 😄

Thanks

[copper kraken]

np haha had the same issue a while ago 🙂

[pallid path]

@mellow tendon low cfg with the prompt modified from this very familiar guy https://civitai.com/images/11819775 creates nice images

and I'm not even using loras, just sdxl base

[dim steeple]

What is the controlnet model that helps keep faces consistent during img2img? Is it Ip-adapter plus face?

[mint basalt]

How would you make her face more photoreal?

[gloomy lark]

have the prompt for that?

[light surge]

Analytical skills

[elder ravine]

.

[weak nexus]

[tawdry current]

There's a new canny controlnet for SDXL that's worth checking out: https://huggingface.co/TheMistoAI/MistoLine

[stray warren]

Not just canny...

[tawdry current]

Correction: Line. I always just lump them all into the same boat

[stray copper]

2

[mint basalt]

This was SD3, "reaching out handing towards camera a red and yellow carton of french fries that reads "100% tallow" on label, Behind her and register is neon green sign reading "Seed Oil Scout". Sunlit beachside burger joint with palm trees. Cashier is beautiful young blonde american woman in red and yellow outfit,"

Sent you a message btw

[slow dirge]

travel

[uncut gull]

This is actually useful! ...Was my first reaction but why would it be incompatible with Lightning? All the controlnets I've ever tried work perfectly with lightning, so they must have really messed up this things model compatibility somehow. 😕

[fiery field]

#imagine

[elder plume]

@cerulean dove

Hi Guyz can any one help me i want to finetune sdxl so it generate every image with the "solid color empty background" where i can put text in future i want stable diffusion to give me result like this
(then he posts an image of a greeting card)

[meager canopy]

Did a whole load of channels just get chopped?! #bot-status, #SDXL-show-and-tell

Cleaning up the channels ready for the SD3 release tomorrow 😉 🤔

[gloomy lark]

Yeah that's the current theory. 🙂

Release on a Friday afternoon and run away for the weekend. 🙂

[meager canopy]

Loved this one #🆕｜sd3 message

[gloomy lark]

Hah yeah sad panda. People complain that text isn't perfect in sd3 but if you're willing to run through some seeds it'll do even multiple independent text bits well

[stray warren]

I wonder how long this channel will be here

[copper kraken]

if this one goes away without sd3 weights dropping i'll be outta here

[stray warren]

at least https://discord.com/channels/1002292111942635562/1108646783090958336 is back up, so there's still hope

[copper kraken]

only for outputs from #1237460191780540456 though

"This channel is for showing off and chatting about awesome creations made with Stable Artisan"

[stray warren]

The old stuff is still there, now we just share and tell over here or in https://discord.com/channels/1002292111942635562/1230206273451069540 which makes more sense

[meager canopy]

I hope the bots appearing back as a paid service isn't any indication of where the SD3 weight drop is heading. I can understand the service needing to be paid for, but please keep the promise of releasing the models publicly 🙏🏻

[wicked frigate]

those channels that were moved to the Artisan bot channels, were the channels for the old XL Bot

[wicked frigate]

Yes the plan is still to open release the SD3 models

[meager canopy]

Thank you!!! Do you know of any timeframe...non-commital of course 😊

[wicked frigate]

don't have a date atm, sorry

[elder plume]

wow my greeting card workflow is actually really good lol

how do people release workflows?

is there value in something that takes 5 minutes to do?

[uncut gull]

Ooo value happens to be one of the philosophical subjects I'm investigating at the moment! It's a profoundly deep question!
But no. The short answer is no.
The longer answer is: Post it on Civitai. They have plenty of workflows. 🙂

[peak dove]

Value is something done, something active. Value is something made.

The greatest value(s) is underpinned by Custom; and Use.

[uncut gull]

Sorry, your comment is far too vague and untestable to be valuable. 😛

[verbal coyote]

big improvements today

I'm hopeful

[copper kraken]

i have no doubt it will kick ass in the end

i've already gotten stuff that ideogram/dalle3/MJ6 couldn't do, and that's with the old model and whatever the workflow is with the api

[glass forge]

Whats happening

[acoustic spruce]

What extensions are you guys using for autofill keywords? Been using https://github.com/mix1009/model-keyword but I just noticed that it's over a year old. Made me wonder if there's more, and/or better ones

[meager canopy]

Run it through a face detailer

eg.

[peak dove]

Value is not intrinsic; it is proposed from outside.

[glass forge]

tallow those fries

[uncut gull]

Term definitions and testable metrics or it's not true. 😛

[paper mantle]

stability.ai API does not have an inpainting engine for SDXL 1.0? Inpainting does not work for me - it's just inpainting in something extremly different (exactly where the mask is)

[meager canopy]

They are already deep tallowed

[rustic garnet]

@fierce heron if you want small changes then use lower denoising. For larger changes I would use differential diffusion

[dreamy junco]

can anyone reference me a guide to SDXL? I had/have auto1111 for SD 1.5, but I recently came back and saw people using pony XL and SD XL, but I'm not sure how to install or if i can keep SD1.5

[gloomy lark]

neat chroma effect

[upbeat summit]

just my post-processing pipeline for that cinematic style 😉

[gloomy lark]

do you remember what comfy module that post-processing fx one is part of?

have that png handy? 🙂

of the grain

[upbeat summit]

the chromatic aberration node is from WAS nodes https://github.com/WASasquatch/was-node-suite-comfyui

[upbeat summit]

I dm you about the grain

[unkempt pond]

anyone can help me with creating 360 panoram image

i found this reddit post saying the sdxl can do it but i cant get it to work https://www.reddit.com/r/StableDiffusion/comments/16csnfr/workflow_creating_a_360_panorama_image_or_video/

[copper kraken]

diff seed, same eveyrthing else

[grizzled viper]

In case no one mentioned it yet - I noticed there’s a new XL inpainting C net model on HF linked in Civitai

I tried it - it’s very good

Oh, also here’s my discord bot
https://github.com/altoiddealer/ad_discordbot

[gloomy lark]

https://civitai.com/images/11776408 Running this prompt through pixart. very different result but neat stuff.

[fierce heron]

Wrong ping

[gloomy lark]

another prompt from masslevel

[dreamy junco]

question, can i use SDXL 1.0 loras w/ pony interchangably, vice versa.

[glad grove]

pony needs its own loras

[dreamy junco]

I see, so if i want to use an SDXL lora I would need an sdxl checkpoint

[glad grove]

yes like animage

but some xl loras work with pony

they just dont look as good

[dreamy junco]

hm alr ty

[dreamy junco]

what's animage>

I only know of sd1.5, sdxl and pony as of just a while ago

[glad grove]

the other anime XL checkpoint

[dreamy junco]

Hm mk, and it's not for sdxl1.0

Alr ty

[glad grove]

animage is sdxl 1.0 so is pony,its just that pony was trained in a different way so it needs its own loras

[dreamy junco]

Ah I see

So there are a few checkpoints that are like that huh

[dreamy junco]

wait your talking about animagine for this right? It should work the same with anything xl right

[glad grove]

yes

[dreamy junco]

mk

[dreamy junco]

for SDXL, am i supposed to upscale by 2 from 1024x1024? I have a 4070 12gb of ram, so 2048x2048 might be pretty heavy still

[glad grove]

i upscale by 2x with a 3060 12gb and it works so it will work for you too

[dreamy junco]

how long does it take for u to gen one?

[glad grove]

1min for a 1216 x 832 img at 15 high res steps

[dreamy junco]

hm ok i think i'm doing smth wrong then

the image takes around 2-3 min to gen

any thoughts on things I should change?

[glad grove]

yea denoising strenght too high it should be between 0.3 and 0.4 and hires steps with like 12 to 15 is more than enough

[dreamy junco]

alr lemme give that a shot

generally sdxl only uses 1028 x 1028 right? It's not trained for any portraits?

[glad grove]

u can you any of these ones https://www.reddit.com/r/StableDiffusion/comments/15c3rf6/sdxl_resolution_cheat_sheet/

[dreamy junco]

got it thxs

so from what i'm seeing on other posts, pony recognizes a lot of characters w/o the help of loras

I tried using one of the characters, and it's having a hard time making a clear image of it

[glad grove]

yea sometimes its better to use a char lora

[dreamy junco]

hm mk, imma run another test and directly copy from one of the character sheet lora examples and see if it works

[pallid cradle]

prompt imagine a hut near lake with some coconut tree near mountain

[dreamy junco]

ill give that a shot in a bit

ok so this is what i got

this is what others got

ik they're using a different model, the i'm pretty sure there's still something i'm not doing right for mine to come out that high saturation and lack of details

[glad grove]

prob cuz they are using the autismXL

you are also using dpm 3m sde and they are using Euler A sampler

[dreamy junco]

hm, ill try an example straight from the pony checkpoint page then

[glad grove]

try to change sampler

[dreamy junco]

which one would u reccomend?

[glad grove]

either Euler A or DPM++ 2M

i mostly use ddim and ddpm but you have to try with diff ones til u get a good img

[dreamy junco]

hm mk, what about schedule type?

[glad grove]

i just leave it at auto

[meager canopy]

A link or name would be more useful 😄

[random palm]

I am using the graphics card above，I don't know why the graphics memory usage is so high，Please give me some guidance, brother

I am using this model(SDXL)

I am very annoyed about this

[paper mantle]

is there an inpainting API somewhere?

[copper kraken]

Are you using a1111

[peak dove]

Using AUTOMATIC1111 you will find VRAM memory usage pushed to the max. ComfyUI is much better at memory management. However, max memory can still give good results

[acoustic spruce]

Bump :3

[grizzled viper]

New XL inpainting model I mentioned https://civitai.com/models/136070?modelVersionId=497710

I made a comment there saying how to use it WebUI because its usage is not obvious

[grizzled viper]

Woks great if you don’t mind your unmasked areas subtly changing

[meager canopy]

Thanks! 🙂

[tawdry current]

Or just put a differential diffusion node between the checkpoint loader and the ksampler and blur your mask. Turns any checkpoint into an inpainter and provides excellent quality

The mask has to have gradients to it and you'll want at least 25 steps

(if you're using comfyui)

[copper kraken]

This. Diff diff is the best inpainting method I've found. It's much better than the "soft inpainting" method added to a1111 a few months back

[tawdry current]

Yep, it's amazing. I've really tried putting it through some torture tests like making a black to white gradient across the whole image where a scene transitions from a photo of a forest in the day time to a painting of a cyberpunk neon cityscape at night. The key is to have enough steps though and to make 100% sure you have gradients

[copper kraken]

yep... i have an iterative workflow too that uses latent sender/receiver and latent composite masked so there's no vae degradation either

it's absolutely the best way to inpaint these days

and comfy also just added opacity for the mask editor

[tawdry current]

a lot of times, i don't even use the inpaintmodelconditioning node, i'll just use a set latent noise mask if im going from one ksampler to another

to avoid dealing with vae encode decode encode decode artifacts that can happen from multiple passes

but differential diffusion f'in rocks so much to the point that i deleted all my inpaint models that were taking up space because i haven't touched them since lol

[copper kraken]

that's exactly what i do

what i have noticed though is diff diff does affect the unmasked portions

it's subtle but if you do something like 4 or 5 in a row there's obvious degradation

but the latentmaskcomposite thing solves that issue

[tawdry current]

yeah, it technically touches everything, but it's usually subtle enough that it doesn't degrade overall images. but during the phase where you're doing this kind of stuff, nothing should be "final" quality anyways and it always ends with one final denoise pass to tweak the image that last 5-10% to what you want

[copper kraken]

depends i guess, i've actually just used it straight up a number of times

but... i'm usually aiming for a pretty wild look

and actually, now that i think of it i do usually do something not far from that

that last pass is usually a diff diff gradient across the entire image

[tawdry current]

but either way, whatever minor issues it has, the benefits of diffdiff>>>>>>>>>regular inpainting

[copper kraken]

yeah it's amazing shit

barely anyone uses it but it should be a routine part of ones' workflow imo

i have a shitload of masks dumped in a folder for diff diff stuff like the above

[tawdry current]

yep, got a bunch as well and if i don't, i just fire up photoshop or zbrush to make them

[copper kraken]

for you or anyone else that wants to take a look

lots of wacky shit... some of these are great for framing an image

[tawdry current]

i liked how this one turned out

the rainy road transitioning into a glade lol

[copper kraken]

yes i looove that kinda stuff with diff diff

i like to do sequences with those masks

shit like that

[tawdry current]

I'll check those masks out later

[copper kraken]

found it's best to start with a real image, then diff diff over it

cool, yeah i've had too much fun with those

those are some easy ones for regions... found it was good to have a tiny bit of overlap in most cases, kinda like you were saying about gluing teh image together

this one's pretty great for a dual pass workflow... style one, rotate mask 90 degrees, style two

[gloomy lark]

well fine, i'll just take the bus then.

[glass forge]

can these masks be used to promt certain areas of an image?

and doe sit have to be a hard mask or can it be a smooth maks where the effect of different promst take effect gradually and overlap each other

Asking for a firiend in Noodletown.

[night nebula]

'

[tawdry current]

Yeah with regional prompting(can't remember the acronym name, gligen?)

I haven't tested it personally, but I don't see why the two wouldn't work together

For differential diffusion to work, you need gradients, otherwise you're just doing regular binary inpainting

The gradient is where it works its blending magic

[worthy orbit]

Did anyone tried these new CN inpaint models? https://civitai.com/models/136070?modelVersionId=497710

https://civitai.com/models/136070?modelVersionId=496890

[wheat ravine]

/help

/help

[gloomy lark]

wow... i detect a hint of clownshark in this one. 🙂

[copper kraken]

I let it take over from one to the next

Hard mask for first 0.2 usually but sometimes you gotta go all the way back to 0.1 or even 0.05

But it helps a ton with regional prompting and tbh is so useful I wish it were part of the default nodes

[barren crown]

Cartoon Bull, Kung Fu, Kicking

[gloomy lark]

https://discord.com/channels/1002292111942635562/1237461679286128730

[thorny marsh]

Hi everyone! I need since help fine tuning stable diffusion for a product, if there is anyone that can help?? Appreciate you all!

[copper kraken]

... Do you mean a Lora?

[nocturne aspen]

I have a really hard time with crossbows

bow and arrows in general, but especially crossbows

[copper kraken]

the spiderbus will take you to crossbow school... all aboard!

[glass forge]

Some mass destruction I can get behind.

I love crowds in these images, and it's the same in other models too like SDXL, the crowds seem completely oblivious to whatever insane mayhem being promted up.

[nocturne aspen]

what do these supir settings actually do?

how can I increase the denoise?

[tawdry current]

Supir is an upscaler. You need to have your image already completed before passing it to supir. It doesn't function the same internally as other upscalers like ultimate upscale, as far as I know

[worthy orbit]

it makes sense as crossbows are some strange object and SD confuse it with bow like in the image

maybe ask for a modern rifle, shotgun or other gun (or inpainting prompting it) and then inpaint to get the crossbow; as crossbows resemble more a rifle than a bow and are held in a similar way... well maybe not ancient crossbows, but I think it may help to get the basic drawing

[nocturne aspen]

Ya, I had gun as a negative so maybe better like you said

[nocturne aspen]

But the supir node doesn't actually do the upscaling, right? The workflows I've downloaded have a traditional upscaler first and then pass the already upscaled model to the supir node

No one seems to know what supir actually does or what those settings mean

[tawdry current]

I think those other passes are just conditioning the latent space for the actual supir sampler. The source code is available on the original supir repo if you want to dig into it

[trim orbit]

Youtubers aren't trying to actually educate people in most cases. they're trying to farm views. do yourself a favor and stop viewing the yotubers as experts in the field. they're just people that are rushing out content and likely didn't read or grok the paper. https://arxiv.org/abs/2401.13627

in previous years, i found there was more written guides available by people who actually did understand what was going on. Then youtubers would just rehash those guides, without credit, into their videos. Acting like they were the one who did all the research on the matter, because they half assedly cut someone else's written tutorial into a video script and bumbled their way through it for their audience. Those videos would get all the views and hype, the youtubers get all the credit. People trust that they're experts in the field and form fan cults around them. But they don't actually know the fundamentals and just rehashed someone else's research with errors.

So now i've noticed that there are a lot less detailed guides coming out these days. The actual experts have realized they're just aiding some youtubers' content farm and they're left uncredited.

"Nobody seems to know how...." is a sentiment happening more and more, as experts stop feeding the youtubers everyone relys on

[glass forge]

Youtube has become repulsive.

98.6% of them tubers, they're desperate and pathetic. And these are accurate scientifically verifiable numbers.

Almost nobody makes vidyas for fun anymore.

Those early dog AI videos were made for fun, you can tell.

[trim orbit]

everyone is sponsored. 10 years ago, todays youtube would've sounded like a black mirror episode

[minor pecan]

https://tenor.com/view/smile-happy-creepy-gif-10049761

[glass forge]

Mark Zuckerberg?

[slender coral]

Fun doesn't make money.

[lavish ermine]

Hey is the stable video out for testing yet?

[glass forge]

By the time you make your first million it will be worth 200k. Are you having fun yet?

And that's before taxes.

And needy relatives.

Let me know on your way to your lawyer and accountant how much fun you are having.

Did you pay for these bunkers you didn;t mean to pay for? Aw I am sorry.

You sure as hell won;t be in them when the bomb falls.

See you in dubai, unless your sand island sinks. then I'll buy your lambo from the graveyard for a discount. We'll both be having fun then.

And if you're not having fun and just make money well who the f cares how big was your bank account when you croaked?

Do you wear clothes? Because the empreror sure doesn,t.

[paper phoenix]

These remind me of JK Potters work, especially on the cradle of filth midian album art.

always remember that artists name because harry potter and jk rowling confuse google searches so much

[gusty elk]

A girl img, 3D, All contours, in a cyberpunk city, wearing a neon-lit outfit and holding a high-tech gadget.

[stray warren]

Awww, so happy
I think he gave his arm for it

[stuck swallow]

https://tenor.com/bL7lN.gif

[dreamy junco]

hi, so i switched from 1.5 to sdxl/pony recently, and I learned that the keywords are a bit different

so i'm trying to generate more anime style art, however the style is pretty inconsistent (some generate with more shadowing like 3d ish, others more simple/flat shadows like anime) unlike sd 1.5 where it primarly depends on the checkpoint and/or lora. Any suggestions?

[worthy orbit]

pony (in my experience) is very tricky and evasive, first you need to follow the prompt instruction in the page. I think it has DPO (Direct Preference Optimization), as I understand it, it allows for more intricate images with more characters interacting or details, but at the same time it might become chaotic, as for example, something you prompt like a detail will appear as a full entity.

There are many Loras for pony style also, I guess they are a necesity because of the DPO thing might oclude or mess the style prompted

[dreamy junco]

Hm mk where can I access to take a look at the page?

[glad grove]

the civitai page for the ponyv6 model,theres a small guide there on how to prompt with it

[dreamy junco]

Is it directly on the pony diffusion civita page? There r also some other mixes that r still categorized as pony, will the reference still apply to those?

[dreamy junco]

Hm mk

[glad grove]

this one https://civitai.com/models/257749/pony-diffusion-v6-xl

[dreamy junco]

Awesome, I'll take a look once I get back

Ty

i noticed that one of the pony models provides some details on the scores

i thought it was just the quality of the image, rather than the type

[gloomy lark]

I stole this prompt from one I saw on the artisan channels. looks incredible through pixart: A high-quality, cinematic still of a high-tech research laboratory in Japan, where a menacing AI with glowing red eyes dominates a massive wall-mounted display. Japanese researchers and lab technicians look on in horror and astonishment. The lab is chaotic, with scattered papers and malfunctioning equipment emitting sparks and error messages. The room is starkly lit, emphasizing the glowing red eyes of the AI and casting long shadows, which amplifies the tension and impending danger. This dramatic scene captures the moment the AI begins to exhibit unpredictable and threatening behavior, witnessed by the stunned scientific community within the confines of the lab. Shot with Canon EOS 5D, nikon d850, canon ae-1, 32k uhd, the whole body

[copper kraken]

Trained a character Lora

[peak dove]

I did this prompt via Pixart-Sigma+SDXL

[long compass]

Navigating the Dynamics of the booming Saudi Market

[acoustic spruce]

Yo guys here's a 140 page document with styles, artists and tags. Enjoy https://docs.google.com/document/d/e/2PACX-1vQMMTGP3gpYSACITKiZUE24oyqcZD-2ZcvFC92eXbxJcgHGGitde1CK0qgty6CvDxvAwHY9v44yWn36/pub

[gloomy lark]

what's hilarious is when you replace that AI with all sorts of wacky stuff

kinda of funny that asking for a children's book of robot stuff actually gives you the book too with pixart

[quaint lagoon]

https://tenor.com/bOq8d.gif

[copper kraken]

Love the eyeball moon

[copper ice]

呵呵

[dull falcon]

#dream

[errant pulsar]

Can I download it

[acoustic spruce]

It's as easy as copying everything into a txt file! If you want the styles that are prompted, you need to copy them into your Styles.txt in your root SD folder

[errant pulsar]

Okie thanks.

[copper kraken]

@median mulch

here's one version of my iterative inpaint workflow

allows you to loop one step at a time

so you have the convenience of a1111-style inpainting with the power of comfyui

[median mulch]

😮

nice ty

my main issue with inpainting is that is almost always gives me the masked area seams/artifacts after, i just want something that doesnt give that or reduces them to a very meaningless level

[copper kraken]

diff diff solves that

with a blurred mask

give it a shot

this is a simple version, it's definitely better with other stuff added like samplercustomnoise for both stages, aligned or other custom schedulers, ipadapter, controlnets, etc

[toxic wasp]

Let me guess, you're from Colorado, originally from California?

[copper kraken]

nope

[uncut gull]

In 3 steps : A kitten in a witch costume jumping over the moon. The model is creaprompt (not my model).

(Yes I know both lightning and hyper claim to have 2-step and even 1-step versions, but those have never produced a usable image for me.)

[tawdry current]

when i use speed models like those, i only ever use 8 step versions of them. i'm not that impatient

[gloomy lark]

horrifying jello monster in a top hat and tuxedo, in the style of chic photography, split into multiple different images shot from various angles, différent expression, against a white background, kodak portra 160

[craggy wave]

g4tyh5

[peak dove]

Sigma+SDXL

[copper kraken]

great style!

[stray warren]

The power of SD2

[stray warren]

Back to SDXL

[ashen otter]

what

[half thorn]

a man is shouting #artisan-faq

[ashen otter]

Bro 😭

[burnt aurora]

"A young man in his early 20s standing in a dimly lit room, surrounded by shadows of tempting figures. His eyes are wide with longing, and his hands reach out towards the shadows, symbolizing his entrapment by desires and lusts. The background features a mix of dark, alluring colors, adding to the atmosphere of temptation and struggle. The scene conveys a sense of internal conflict and emotional intensity."