✨｜sdxl

[kindred shoal]

this thing crashes every other prompt

i dont understand why it keeps doing this with sdxl.

i have to restart forge in order to get it back to normal.

[glad grove]

yes i get that error sometimes on forge XL

[kindred shoal]

i dont know why it happens though.

it's getting on my nerves

...it's fucking cutoff

i turned it off and it worked

[glad grove]

yea some extensions can trigger that error

in my case it was CADS

[kindred shoal]

i kinda need cutoff though.

i have so many colors that it won't work otherwise.

[tawdry current]

this is why sd1.5 prompting for sdxl isn't a good idea

[glad grove]

its prob not cutoff because i use it and doesnt happen,it happens when u enable an extension gen some imgs then disable that extension and enable it again

[glad grove]

pony uses its own tags plus the booru tags because it was trained that way,it would be a bad idea not to use those tags

[tawdry current]

it's still based off sdxl though, right?

and does it still use clipg/l?

[glad grove]

no pony uses booru tags,it was trained that way

just like animagine

[tawdry current]

if it is still sdxl at the core, as in trained from the base model, then the clip information is still relevant. afaik, models like pony are just lora merges

[kindred shoal]

yeah, i need cutoff. it's mixing colors way too bad.

[tawdry current]

where they trained a bunch of loras and merged them into a model like the base model

[glad grove]

wrong,pony is not a lora merge lol it was trained with 2.5 million images

[tawdry current]

from 100% scratch?

where they spent hundreds of thousands of dollars renting a server farm to make a whole new base model?

[glad grove]

oh wait a sec you are that guy who was saying the gov was reptilian right

ofc you would say that

[tawdry current]

hell no lol

that's some schizo sounding crap though lol

[kindred shoal]

i like how far this devolved.

[glad grove]

yea i remember that guy its the same one who claimed xl controlnet models were superior to 1.5 ones

[kindred shoal]

dude it's still not working and i restarted forge and immediately turned on cutoff in the new window

[tawdry current]

they are just as good lol.. you're just using them wrong

[glad grove]

ofc you would know that,you are the expert here with millions of models on civitai

[kindred shoal]

..can you not argue and help me out

😭

[glad grove]

ok try to disable the extension on extension list

[kindred shoal]

restart ui, enable it, restar ui again?

[glad grove]

or try to just enable it and then leave it on

yea disable it on the list of extensions then restart ui again

[kindred shoal]

okay so i disabled it, reloaded ui, enabled, reloaded, testing now

[tawdry current]

that's barely even over half strength. i was going to troll around with some husbando crap instead of the typical waifu stuff people always spam. This still has the same prompt from earlier helping him figure out the sheer sleeves thing

[kindred shoal]

it still errored out. i'm gonna try and fresh install it

[glad grove]

do u have any other extension?

[kindred shoal]

these are the others i've installed

[glad grove]

try with an empty prompt maybe it will work

[kindred shoal]

it only messes up when cutoff is turned on...

okay so cutoff is on and empty, and it didn't error out

[glad grove]

now dont disable it and try to gen a few imgs like a batch of 6

[kindred shoal]

should i try putting text inside of cutoff?

[glad grove]

no just leave it empty

[kindred shoal]

alr, it'll take a min to make a batch

i'll make a batch of 3

[glad grove]

if it works try another batch but with a cutoff prompt

[kindred shoal]

with cutoff, you just put in the colors right?

liek red, blue, red, purple etc etc

[glad grove]

it works with other stuff too like butterflies for example

[kindred shoal]

ahhh so does it work better with the full tag?

like white shirt?

[glad grove]

its better with just colors because thats what bleeds onto other stuff,other stuff that tends to bleed is "frills"

but depends on the model

[kindred shoal]

ahh, got it. okay so it ran the batch just fine

with an empty cutoff

error'd out with a filled in prompt

works with a cleared prompt

[glad grove]

yea in my case it fixed itself, was hoping for a fix but forge been dead since 2months

[kindred shoal]

that's so stupid

i'll try and clean install cutoff, because i really need cutoff.

AHA! I FIXED IT

It was embeddings I was using!

[copper kraken]

he made his last commit on github a day or two ago, so he's alive

just to a private repo

[kindred shoal]

i removed the embeddings zPDXL and zPDXL-neg and it worked

zPDXL must use color tags or something

[glad grove]

[native moon]

a rodant!

[kindred shoal]

it does not look like sheer sleeves works here

because it's not a booru tag

see-through sleeves is a booru tag though

[glad grove]

yea this one works see-through_sleeves

[kindred shoal]

i'll try it after this batch. cutoff seems to not be working again unless i have the colors in the wrong order

[minor pecan]

[neon fjord]

Hello
I don't know where to place this question but i tried hard to get decent images from various sdxl checkpoints and workflows in comfyui but still i can't get the image quality which I can generate so easily with SD1.5.
Is this a common issue?
I know this is a very general question but is this normal to struggle so hard with sdxl in the beginning?

[tawdry current]

don't, feed, sdxl, prompts, like, this and type more like this. place things from most to least important and stay under 40 or so words(<75 tokens). your negative prompt doesn't really need much other than maybe nsfw, text, watermark. other than that, there really isn't any need for the mile long sd1.5 negative placebo prompts (most of which are just shifting the prompt seed and similar results could have been obtained by inserting empty tokens like , , , ,). Also, make sure your latent image size is set right like 1024x1024 or some other combination that multiplies to ~1 million pixels. 896x1152 is a pretty good one for portraits. I think 1344x768 is a pretty common one as well for landscapes and is almost equal to a 16:9 ratio (good for wallpapers)

don't set your cfg too high either, stick to somewhere between 4 and 7, the realism models like juggernaut prefer lower cfgs like 3-5

[copper kraken]

i rarely go below 5 myself

tends to get a "dusty" look

[tawdry current]

well the lower you go, the closer you get to the reference dataset and the less your negative prompt works. personally, i stick around 4-6. i always laugh when someone posts some "really good portrait" image and you can pretty much automatically tell they set the cfg to like 1 and you're basically getting an almost pure dataset image that completely ignored their prompt

[copper kraken]

it just starts losng a lot of definition and contrast

at those low values

[tawdry current]

yeah sometimes, but that's also good for if you typically do cleanups in photoshop afterwards. that way, you have more room to play with the darks and lights. depends on the cfg and the sampler a lot too

well and the model in general

[glad grove]

if u talking about anime style pics u need pony XL to get quality similar to 1.5 checkpoints,go to civitai and download it from there also read the guide on the pony page to know how to prompt with it

[vague tapir]

damn what promt did you use for 3 image ?

[vague tapir]

https://tenor.com/view/plague-marine-warhammer40k-chaos-space-marine-death-guard-nurgle-gif-16743287

[shut crescent]

info should be in pic

[vague tapir]

only you can see info

[shut crescent]

PMed u

[vague tapir]

thanks

[vivid rampart]

Does anyone know math function for sdxl prompt weights?

[blissful ruin]

Is there any ways to use XL for free? I've used SD through Google Colab before

[shut crescent]

localy

[blissful ruin]

Okay what if you're broke

[shut crescent]

i dont think it would be possible to use it for free then localy one even collab needs money

[blissful ruin]

But Colab is cheaper than running locally if you just want to run a few gens right

[shut crescent]

then u need to use collab

[blissful ruin]

The colab I was using didn't use SDXL

[shut crescent]

i know that its possible to use your custum models but collab will provide u GPU

[shut crescent]

u need to read documentation

[blissful ruin]

Ty bud

[shut crescent]

try to save some money for own GPU

[shut crescent]

3060 12GB

[blissful ruin]

That all you need?

[shut crescent]

and RAM

[burnt shadow]

examples?

[wet nacelle]

Awesome

[burnt shadow]

you can easily go higher than that, up to 15 from experience

[astral path]

anyone knows a good strategy to fix this? adetailer doesnt fix it, happens a lot for me with people wearing glasses

[wet nacelle]

you could try to just run the seed back through the pormpt with an upscaler.

[ashen otter]

1+1=3 anyway hi

[elfin rock]

If anyone is interested in a free Alternative to Midjourney, I've been working on a bot that can generate images at a similar level if not better. The images generate as fast as 8 seconds. LMK if you want to test it out, we are looking for feedback and suggestions from anyone who is willing to help out.

[cyan crown]

https://civitai.com/models/369380?modelVersionId=412708

[cyan crown]

is it a model?

[half roost]

Life of an artist

https://tenor.com/bSmq0.gif

A digital illustration of a steampunk library with clockwork machines, 4k, detailed, trending in artstation, fantasy vivid colors

[minor pecan]

Heres are the images you requested.

[elfin rock]

WDYM?

it is a discord bot that uses a SDXL model for txt2img

[gloomy lark]

Dude you are the GOAT. I use 25% cute 3d render lora on all my merge models.

now this...

[cyan crown]

thank you

for this I suggest to use with Realvis lightning 4.0

realvisxl 4.0 lightning

[half roost]

A model wearing a black off-the-shoulder dress, pearl earrings and a small gold chain with a delicate flower pendant around her neck. The background is a solid color and simple, highlighting her profile. She has blue eyes and long hair in loose waves, giving an elegant feel to the overall look of the outfit. Her expression radiates confidence as she poses for the camera, showcasing a stylish and fashionable style in the style of a fashionable portrait., vertical aspect ratio

[gloomy lark]

awesome stuff

[cyan crown]

thanks

[gloomy lark]

so one question though. what's the possibility of removing the need for the trigger word? is that something I can do myself in any way?

I like using that cute3drender with a merged model for just an overall look to everything.

I see that it doesn't particularly kick in without "lego world"

[cyan crown]

for me it works without trigger

but the trigger helps sometime

[gloomy lark]

so this is the same prompt and seed, but without lego world. the look isn't complete like it is with the trigger

disregard the hands on chest, that's probably because of the detailed tattoos.

[cyan crown]

yes with trigger works better

can you give me the prompt?

[gloomy lark]

a stone giant wrapped in chains and tattoos is crouching in front of a gate looking down at a small group of diminutive people who are trying to gain access to the cloud town.

[cyan crown]

this is what I got

[gloomy lark]

i wonder if that's where that model comes in.

@cyan crown ok, i put the strength to 1.5 and now it's very consistent even without the trigger word

on my model

[cyan crown]

BTW it seems to work well

[gloomy lark]

yeah this is great stuff.

[gloomy lark]

final value ended up being 1.25

the lego hands really solidified with the new aligned steps nvidia scheduling as well

[glass forge]

Why is every video abouty ELLA chinese...

[tawdry current]

Because the creators of it are Chinese and because China has well over a billion people?

A lot of the AI advancements have been driven by Chinese researchers

[gloomy lark]

I dunno about you guys, but that aligned steps sigma stuff fixed guns for me. They're all mostly correct all of a sudden.

[upbeat summit]

it can fix things for sure 🙂

[wet nacelle]

And sometimes German

[minor pecan]

robutussy

[scarlet sparrow]

"Design a vibrant and avant-garde sticker featuring an asymmetric layout with dynamic, irregular geometric shapes, such as fragmented lightning bolts and stars. The primary colors should be bright blue, vivid green, and shiny silver, contrasting against a glossy metallic silver background. This sticker is intended to be applied to a cylindrical surface."

[copper kraken]

Here is the image you requested.

[inland hound]

[restive garden]

A colorful birthday celebration with balloons, streamers, and a delicious cake. The cake has the words "Happy Birthday Sharvi" written on it in frosting.

[neat crest]

Hey guys im using Stable Diffusion XL Turbo on ComfyUI and these are the images im seeing how do i fix this

[hardy marsh]

画一个哥斯拉

"Design a vibrant and avant-garde sticker featuring an asymmetric layout with dynamic, irregular geometric shapes, such as fragmented lightning bolts and stars. The primary colors should be bright blue, vivid green, and shiny silver, contrasting against a glossy metallic silver background. This sticker is intended to be applied to a cylindrical surface."

[tawdry current]

show the workflow, you're probably not using the right scheduler settings

[neat crest]

Here it is

[tawdry current]

euler ancestral and a cfg of 1

and one step

[neat crest]

That worked even with 4 thanks a lot

[tawdry current]

from the comfy examples

ahh good, np

if youre doing 4 steps, just switch to a lightning model, they are better in pretty much every way and do things at 1024 instead of 512

and you dont need to use the custom sampler with them either

[lapis cedar]

Can anyone send me a link for SDXL controlnet? I had to use SD 1.5 because I couldn't find SDXL.

[meager canopy]

https://huggingface.co/

[lapis cedar]

I meant a link for the controlnet.

[meager canopy]

Which "controlnet"?

[lapis cedar]

let me show you sd 1.5

[meager canopy]

Do you mean the model? There is a search box on that site.

[lapis cedar]

https://huggingface.co/lllyasviel/ControlNet-v1-1/tree/main I used this one: control_v11e_sd15_ip2p.pth, is there the one for SDXL?

https://huggingface.co/ckpt/controlnet-sdxl-1.0/tree/mainW which one of these is a similar one but for SDXL

I tried a few of them but it wouldn't process it and instead stop at that that point.

[meager canopy]

Just search "controlnet sdxl" and choose the one you want

[lapis cedar]

ok. thank you

[lapis cedar]

also any recommended hand fixer models for SDXL?

[meager canopy]

https://huggingface.co/Bingsu/adetailer/tree/main

[lapis cedar]

thank you, I'll try it

[shut crescent]

testing wow lora

[hallow barn]

https://tenor.com/1ElA.gif

[wet nacelle]

@humble capeDid you change the sampler in fooocus? I did. I find that default brings in more candid details whereas euler-a makes the scene look fake.

[noble shoal]

From the series: Inconvenient living

[wet nacelle]

ha!

[humble cape]

I not changed the samplers

[wet nacelle]

interesting

[humble cape]

@wet nacelle if you change the sampler you get better results ?

[wet nacelle]

I mean in theory yes. Chance.

[humble cape]

Do you run fooocus localy ? What is your gpu and how many vram , i plan to buy a new pc

[wet nacelle]

Oh okay. Yes I am running local.

I have a 3070TI with 8gb vram.

[humble cape]

8gb is enough ?

[noble shoal]

I would recommend >=12gb. NVIDIA, obviously.

[wet nacelle]

4gb is for fooocus

[humble cape]

Ha okay thanks , i will buy an rtx 3060 12 gb

For image generation and game dev

[mellow tendon]

"Some" skin on show.

[glass forge]

Hoptimus Prime.

[glass forge]

Yes but fooocus waters things down under the hood to achieve such a friendly VRAM situations. The same generations with the same settings on Forge yield far superior results so be aware.

[wet nacelle]

This is also fact

FUCK AWF YOU FUCKER!!! FUCK WINDOWS!!!

[uncut gull]

Hey everyone. I'm looking for a cloud API for Comfy. Does anyone know of one available?
Not something like a vm paid by the hour, just a cloud API hosting Comfy that's always available where you pay for usage. (I remember seeing something like it in some video ages ago, but no idea how to find it now.)

[ashen otter]

Ancient tech bruh

[uncut gull]

Comfy is ancient???

[ashen otter]

no no I just needed to say something random in here

sorry

[uncut gull]

Oh. Well if you know of a cloud API for Comfy I really need one. 🙏

A1111 / forge etc. will do in a pinch though.

[ashen otter]

sorry I don’t 😔

[glass forge]

how do you guys determine when a model is obsolete, when you have 50 6gig models laying around its becomign a problrem

hard to let go coz eahc model has strenghts

:/

[uncut gull]

Back them up onto an old external harddrive, label them "SD Models to Keep", and then put them on a shelf.
Fond memories a few years from now when you dust it off and find the old label. 🙂

You know, back when SD first came out, I was keeping every good image. It took a lot of careful prompting and re-rolling to get good images. It would be crazy to delete them. ...But then...

[glass forge]

I feel ya

also the images take up some gigs

but the damage is nowhere near as bad as the models

esp SDXL

which remind sme

how big is SD3? :0

i have 250 gigs of models

[uncut gull]

But really, what models do you have that you regularly use? I use Proton for prompt-following, CrystalClear's two latest for lora / controlnet compatibility, and Juggernaut Lightning for instant generation. I have a dozen others on my disk that I downloaded to test, and couldn't give me anything special vs. the others.

[glass forge]

good thing i learne dhow to link everythign to one folder (A1111)

durign ym first foray into this i didnt know you could do that

lol

disaster

yeah i only use maybe 5 regularily

more realistic stuff

[uncut gull]

The only other really "special" model I have is dreamlike photoreal, which can do photo-realistic fabric textures. But it's 1.5 and getting it to actually give me what I want is a huge pain.

I wonder what SD3 will replace... Probably not Lightning. Definitely Proton. Hmm.

[glass forge]

chacolrealponyworldm_v10 right ow is probably the most realistic sdxl model

i like the get the ones that come out like last week

the more time in the oven the more refined lol

its amazing how far sdxl came in 6 months

i remmeber first models

so i can just imagine what refined SD3 will be like

[uncut gull]

Pasted into search and found nothing.

[uncut gull]

I use SD for artwork with ControlNet. (And I'm working on a painting app with it built in.) My priorities are controlability and prompt-adherence.

[glass forge]

well it is pony so fighting the nudity is a pain

[uncut gull]

But what's the actual name?

[glass forge]

https://civitai.com/models/401725/chacolrealponyworldmixxl

https://civitai.com/models/388607/realityfuse-xl

[uncut gull]

That looks very... focused.

[glass forge]

as for plain sdxl realityfuse is very realistic

[uncut gull]

What is a pony model?

[glass forge]

idk

sdxl thats more focused on nudity?

the models do great anatomy but nothign else

if you try to set your image to be anywhere other than a bderoom the backgrounds ar egarbage

with ponies

[uncut gull]

Let me see if it can do hands at least then. Why ponies?

[glass forge]

no clue

try them lol

put nudity in the negatives

to tone them down

[uncut gull]

Nevermind it says it will take 2 hours to download? Why? Civitai has never been slow for me before. Always like ~1 min downloads for SDXL.

[glass forge]

well

its the exact same size

as a sdxl

6.6 gigs

[uncut gull]

Yep. Same with all models, VPN rerolled and off. Huh.

[glass forge]

i hate going to civitai i always find new stuff i wanna try

i wa ssposed to archive models!

[uncut gull]

Oh well. I'll let the download manager run in the background and try them later. RealityFuse does look like great skin and fur texture.

[glass forge]

This makes perfect sense.

[gloomy lark]

we heard you like wings, so we put wings...

[glass forge]

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

[glass forge]

Soon SDXL will be so fast we will actually run at -30 steps. As in we can go backward in time while generating. =0

The time machine will be built out of Nvidias not Deloreans.

[ashen otter]

Nvelorean

[gloomy lark]

sigh. ok now i'm doing these in 8 seconds.

[glass forge]

cats for days

[gloomy lark]

pretty much.

[gloomy lark]

hah

[glossy wave]

I love this! Big occasion

[meager canopy]

https://tenor.com/view/ashlabeth-3d-glasses-gif-13311384

[tawdry current]

Been testing out an overkill combo of the align your steps(AYS)+AutomaticCFG+perturbed attention guidance(PAG). It got a bit confused though still with the various colors in the prompt for the instrumentation and the walls, but it still produced some pretty neat results. I should test some with sd3.
"in the style of retrofuturism, an advanced laboratory with pale {red|yellow|orange} analog instrumentation, analog screens, retro tech, pale {duck egg|coral|pea} pastel colored walls, (liminal:1.15), 60s film aesthetic" (I generated a bunch and it randomly picks one of the things inside the curly brackets, so like red and pea or yellow and coral, etc)

[lapis cedar]

Does anyone know a model that makes a similar style to Chinese Donghua? (CG anime)

[wheat tulip]

https://civitai.com/models/422821

[lapis cedar]

thank you

[wheat tulip]

sorry, i wasnt orioginally rtepsonding to you, (didint see ur message above mine) but np anyway

[lapis cedar]

I'll try looking more into these models. I really liked anime CG when I first watched it.

[burnt shadow]

@willow stone

[willow stone]

you use this art style?

what i send

[burnt shadow]

yep

[willow stone]

nice

need color explosion

[burnt shadow]

@willow stone

[willow stone]

terminator 2

moofi try this style

[willow stone]

worms

[native knot]

Cheetos

[willow stone]

need beer plane for this

[native knot]

Currently working on ground support.

[willow stone]

looks like a carpet

[glass forge]

T-1001

[native knot]

It's finally built:

[burnt shadow]

🙂

[gloomy lark]

frog-pigeon

[mystic fable]

Hey all,
Anyone has a solution for sdxl turbo upscaling (without weird ghosting and duplicate things)? the 512 resolution is quite low, but putting anything on top of sdxl turbo messes up the image 😦 no matter what ive tried

[burnt shadow]

how about upscaling externally? Or within comfy?

[mystic fable]

I mean I wanted to automate that step, so I get better quality results straigth away without needing to use another app or service when i get a decent result for a prompt

[burnt shadow]

so you aren´t using Comfy I at least assume?

[mystic fable]

no, I use the webui thing

it looks less intimidating

[burnt shadow]

what webui thing? 😄

[mystic fable]

https://github.com/AUTOMATIC1111/stable-diffusion-webui 😄

[burnt shadow]

well, can´t you upscale separately in A1111?

like in one go?

[burnt shadow]

btw if you like it on the easy side (including upscaling with realESRGAN x4/inpainting/drawing: https://github.com/cmdr2/stable-diffusion-ui

[mystic fable]

sounds good, but in reality its the same thing to launch as the one I use 😄 messing with terminal all the time is what I don't really like with these solutions (I mean how hard it would be to make a launcher that does the same thing in the background)

[burnt shadow]

not quite sure what you are referring to exactly 😀

[kindred shoal]

brooo someone made a chappie lora for xl, that's dope

[gloomy lark]

basically as good as sora

[burnt shadow]

dog food? 😀

[kindred shoal]

n...no

[kindred shoal]

https://civitai.com/models/121549/chappie-xl

[kindred shoal]

..oh

i didnt know that was a thing

[burnt shadow]

🤭

[burnt shadow]

😄

[neat crest]

Guys how do i install SD roop? i did all the steps and installed all the requirments in a1111

[willow stone]

worm gum

[willow stone]

glazgo smile 😃

moofi is for you

[burnt shadow]

thank you, already know him, yet gonna work on it anyway 🙂

[willow stone]

remember i give you teamlab style for me its not works

[burnt shadow]

why not?

[willow stone]

dont know result not same beautiful

[burnt shadow]

Experiment 🙂

[willow stone]

portal

[burnt shadow]

@willow stone working quite well 🙂

[lapis cedar]

Does anyone have experience with DW Pose preprocessor? It doesn't seem to work when I use it in a workflow with a certain images and I don't know why.

[willow stone]

half life

[peak dove]

Some SDXL using ComfyUI

[cyan crown]

https://civitai.com/models/424693?modelVersionId=473171

[gloomy lark]

Fun tip for the day. Just add hyper detailed in front of everything.

[tawdry current]

Hell yeah, I've been on a retrofuturism kick lately. I'm not big on using loras, but I'm definitely trying this one out later

[glass forge]

Abe - "What have you done to America?"

[rustic garnet]

cool!
it's a while since your last model 😁

it's very different from the retro futurism prompt in base model. More detailed/realistic, a little bit like old scifi games

[cyan crown]

yes

[sleek bear]

wow i really love these images! well doen

done*

[wet nacelle]

doen

[sleek bear]

real 😂

also what did you prompt to get this level of realism?

i type in photorealsitic and hd but i cant achive such results lol

[copper kraken]

wrong prompt obv

look at those images... and think of what "photorealistic" usually means: very sharp, crisp, often actually CGI that looks, well, photorealistic

how would you describe those images? probably low quality, cell phone photo, 2000s photography, 90s photography, shit like that

[wet nacelle]

@sleek bearI use fooocus with image prompting.

[worthy torrent]

bro how do you generate image in here

[glass forge]

Put a deformed midget in there.

[wet nacelle]

David Lynch like

[half cedar]

Flying dragon

#🏞｜general-with-images flying dragon

[burnt shadow]

in SDXL?

[gloomy lark]

pixart.

I finally merged my stuff with cosxl. it knows how to do actual true total darkness.

[noble shoal]

Is it now a pixart model or an SD model?

[gloomy lark]

even makes anime colors pop

you merge cosxl with existing sdxl models.

[meager canopy]

Sweet Dreams

[meager canopy]

If anyone isn't able to merge their own model with CosXL, or just wants to try it out, I've uploaded mine here https://civitai.com/models/240590/ultimateblendxl

[rotund stag]

Cat on airplane

[ivory ledge]

落霞与孤鹜齐飞，秋水共长天一色

[prime blade]

A woman stands on the beach.

@rich olive A woman stands on the beach.

[rich olive]

SD3

[prime blade]

A woman stands on the beach.

[hollow frost]

A group of children were playing in the forest and stumbled upon a small ginseng plant, and the ginseng elf was smiling beside it.

[merry thistle]

sea of flower

dream: sea of flower

[shy kelp]

Anyone using sdxl api?

Please tell, i want your help

[shy kelp]

Not so good

[cyan crown]

?

[shy kelp]

This one beautiful

[safe stump]

Hey is it possible to lora sdxl 1.0 into understanding text

Like sd 3

[cyan crown]

SD3 uses another type of diffusion. So no

[safe stump]

Not open source one?

Wow text is good

[cyan crown]

another type of model. There are papers if you search

[safe stump]

No I'm asking is that model open source so i can use it to build something

Sd 3?

1.0 ?

[cyan crown]

SD3 is not yet released

only api

[safe stump]

Api is

How text ?

Is this sdxl 1.0 ?

Can you please dm me the link to the model

[ember mango]

I'd like as much please

[safe stump]

Can you please make a meme with text

Lol this is the first one pop in my fb feed like this for example

@shy kelp

You mean cant make memes like this with that model

Like This

Oh i want to make memes with a ai model

It's just fun i can put my meme ideas into reality

Yeah that's what i thought

But if the model can generate text it can fit the meme well with the font and colors etc

This is really good

Dall e 3 ?

Or sd 3?

Model that you trained ?

No way

That's good evan dall e have some issues with text

[vale eagle]

Wow, I am doing something similar but as a sdxl ft.

[glass forge]

i wish ym dreams were this creative or wer ethis good looking

[gloomy lark]

you need to replace the clip text encoder with a real language model like pixart and ella have done. sd3 does this as well along with giving a new base model and some added abilities.

[burnt shadow]

just fool around 🙂

[burnt shadow]

Smail-Lee´s loRA is simply a trained loRA from what I at least think to know

[minor pecan]

'

[sleek bear]

that's really awesome i liked your art a lot

[sleek bear]

yes you're absolutely right..thanks a lot!

[wet nacelle]

If you call this art then you're insane.

[sleek bear]

AI art*

or lets just say pics 😅

[wet nacelle]

No person that only uses AI can be called an artist. That label goes to the devs.

[wet nacelle]

I agree with this more

[sleek bear]

actually you're right..totally agreed on this

[wet nacelle]

Thankies.

[copper kraken]

...that depends on the definition of "only uses AI"

[wet nacelle]

Stop right now

It makes me furious that you want to do this right now.

You know what I mean. You can't spin my shit.

[copper kraken]

there's no "Spin" there, only clarification, jeez

[minor pecan]

https://tenor.com/view/shrek-painting-paint-art-making-art-gif-22023331

[wet nacelle]

A person that calls themselves an artist when using souly AI. This applies to cruddy sketches inserted into image2image.

[copper kraken]

well, if you have a vision, then you're an artist, maybe just barely in many cases, but yeah

[wet nacelle]

I say no but at this point you are bringing in subjective viewpoints.

I'd day that I'm not being subjective.

[copper kraken]

i'd say i'm being objective, lol

your determination that i'm subjective is in itself subjective 😛

[wet nacelle]

https://tenor.com/view/fuck-merde-the-offic-argh-scream-gif-21560129

[gloomy lark]

look what I drew

def artist: An artist is a person engaged in an activity related to creating art, practicing the arts, or demonstrating an art.

def art: something that is created with imagination and skill and that is beautiful or that expresses important ideas or feelings.

therefore, i am an artist.

noodles are my medium.

[minor pecan]

Art should make you uncomfortable!

Set CFG to 0 to see uncomfortable art

[gloomy lark]

hah

[glass forge]

AI is just a tool m'kay kids?

[noble shoal]

Worse than smoking....

Pretty hard to get off that ai stuff

When SD3?

[random siren]

狗

[glass forge]

May 10?

[meager canopy]

https://tenor.com/view/haha-so-funny-gif-27253208

[lavish shore]

#🏞｜general-with-images a rain day

[vale eagle]

run prompts

[azure dove]

https://tenor.com/bTF3B.gif

Warriors in ancient Chinese armor

[cyan plover]

anybody have tips for the best settings with the lightning or hyper loras?

[crisp owl]

"NO THAT'S THE BLENDER!"

[gloomy lark]

it takes k-cups

[clear spade]

Hi Everyone
Hope you all are doing good!

Is there any way to combine training of two models in single model using automatic1111

[copper kraken]

you're gonna train a checkpoint in a1111?

[clear spade]

Yes, I want to train new concept on earlier trained model using dreambooth.

[copper kraken]

you might want to look into kohya or onetrainer

a1111 is deprecated

[clear spade]

Okay, thanks for the suggestion

[clear spade]

Can we use other extensions like segment anything and controlnet in both kohya or onetrainer?

[copper kraken]

no

[clear spade]

Okay, thanks

[copper kraken]

use comfyui for that stuff

and use onetrainer or kohya to train

[clear spade]

Okay, got it
Thanks a lot!

[copper kraken]

no prob

[zinc cargo]

is the bot down forever?

[copper kraken]

Probably, unless they avoid going bankrupt somehow

[zinc cargo]

make sense

[acoustic spruce]

Random question, you guys seems to be able to upscale your images really good: is anyone doing it through controlnet and/or Ultimate SD upscaler? I've been looking for good guides but I've only found outdated ones from a year ago, or just shitty in general. Would appreciate it a lot if someone wanted to share a new good guide with me 😊

[gloomy lark]

I'm rendering in pixart at 512, then upscaling it to around 1792x1024, similar to dall-e in a single workflow. some people like to go insane with the upscaling, but for images you may only look at once or twice, that novelty wears off. you really don't need tiled upscaling for the res I'm talking about

you could use depth controlnet too, but again, if you're using deterministic samplers, it's not going to get all crazy in these lower resolutions.

[ornate nova]

hey all how do i use stable difusion because i paid for proffessional version

and i only got email and no password and how do i login and make images???

like i feel like lost now

why isnt it simple

[acoustic spruce]

I've also tried to upscale the image right away when creating it. They seem to get good quality that way, but it requires lots of resources which I ran out of sometimes. I have not tried depth, could look into it though!

A1111 Forge

I mean it's basically the same thing as the original UI but with some added settings

So if I understand you correctly, you are using the hires. fix WHILE you are generating the image?

Right, looks similar

Oh you're using 2 at the same time

Yeah it looks way more detailed

Is it possible to use it in img2img?

No problem! I'll check out the extension

Ah alright, thanks:)

That seems like a good idea. Will do!

[safe stump]

Sup

[weak thistle]

hello! I have an application which creates an image using stable diffusion XL pipeline_text2image; then iterates over it using pipeline_image2image.

Things had been working fine for weeks but as of last night I'm now getting an error when using pipeline_image2image, it seems like it is not accepting the previous input which was a PIL image and am now getting this error:

cannot reshape tensor of 0 elements into shape [0, -1, 1, 512] because the unspecified dimension size -1 can be any value and is ambiguous

anybody else have this issue?

[gloomy lark]

can you paste a picture of your workflow?

[weak thistle]

Thanks for the quick response! here is the function i'm using as a wrapper to call SDXL, the pipeline text_to_image part is working ok, it's the image-to-image part that started acting weird as of yesterday

first call to text2image:
original_img_content, image_bytes = hf.get_image_response_SDXL(hf.summarizer(output))

(note that images_bytes is used within the web app and original_img_content was passed to the inpainting call)

2nd call, to image2image
img_content, image_bytes_new = hf.get_image_response_SDXL(prompt=str(hf.summarizer(output) + scenario_prompts[i]), image_path = None, filtered_keywords=inpainting_keywords) #image_path = original_img_content

[gloomy lark]

are you using lightning or something? 2 inference steps? 0 config guidance? these aren't normal numbers that would work.

[weak thistle]

yes, using lightning, the image comes back in a couple seconds

[gloomy lark]

ok, so 2 steps would make sense, but usually the config guidance which is usually 6+ can be as low as 1 for lightning, but not 0.

so I'd start by making that 1.

I'd also verify that the the text2img stage is producing a real image. To rule out that the img2img second stage isn't complaining that I can't make an image of 0 pixels because effectively there's no input image.

[weak thistle]

confirming it does create a real image and display it in the application

[gloomy lark]

what resolution is the image from the first stage?

[weak thistle]

i will try adjusting the guidance as you suggest. it's 512 x 512, about 3.5MB

[gloomy lark]

that's sd 1.5 resolution, sdxl is usually 1024x1024 or some such.

all that said, just from a code perspective, you're saying if there's no result_image, go make one with txt2img and then you're passing that to img2img. But you're not checking that the txt2img stage actually made an image, so you're not doing a second result_image check before throwing it at img2img. but that's a separate programming thing.

[weak thistle]

i'm seeing the result of text2image display in the application so i know that part is working

[gloomy lark]

would be good just to have error logging in case txt2img fails.

sure

you're specifying sd 1.5 checkpoints for both or sdxl?

I'd make sure each stage has the appropriate SD version checkpoint configured.

[weak thistle]

I'm not specifying a checkpoint can you point me a link to what's best practice there? 🙂

[gloomy lark]

ksamplers can complain about tensor sizes, if you're trying to do sdxl stuff but have an sd 1.5 model pointed at it.

well you probably have that configured somewhere in the code there.

if it's already making an image, then at least the first stage has that configured. so i'd make sure the img2img has the same checkpoint configured.

checkpoint = stable diffusion model

[weak thistle]

ah gotcha. i'm loading this when starting up:

pipeline_text2image = AutoPipelineForText2Image.from_pretrained("stabilityai/sdxl-turbo", torch_dtype=torch.float16, variant="fp16")
pipeline_image2image = AutoPipelineForImage2Image.from_pretrained("stabilityai/sdxl-turbo", torch_dtype=torch.float16, variant="fp16")

so that's turbo not lightning 😅

[gloomy lark]

ok that makes more sense why it's 512x512 then.

sdxl turbo is that res, with 1 step, with a cfg of 1.

[weak thistle]

yep i'm just rereading the docs, it says to turn off guidance for SDXL turbo

[gloomy lark]

well, it doesn't like the image its getting for some reason.

so normally, the seed value of -1 is so that it randomizes the seed for each generation. it's complaining that a dimension of -1 isn't valid. it's needs to be betweeen 0 and 512. check your resolution settings and make sure your -1 seed value isn't in one of the resolution fields.

[weak thistle]

ok, good point looking into that

the image resolution is 512 x 512 and it's telling me the size is zero....so.....maybe i should convert this image into a tensor or numpy represetntation?

[gloomy lark]

yeah, at least on this side, that second stage isn't happy about it even though we can see it looks ok. unless that second stage isn't able to load the image file on that path or something.

but i don't have enough info to say one way or another

[weak thistle]

hmm....indeed! well i guess i have decided i'll try to convert to tensor or numpy and see if that does anything . thank you

[weak thistle]

well this is...just....i'm at a loss for words! changing the strength parameter got things to work as they had been

back to default 0.5

😱

now i remember the documentation

the product of num_inference_steps and strength has to be greater than 1

2 * 0.4 = 0.8 hence epic fail

well i'm glad we caught that 😭

but now i found something work doing a github pull request over! as that error message was not very helpful

[icy brook]

anyone on here good at extracting loras?

maybe? I am not getting it to work right. I get a LoRA but it just outputs noise really. I tried with the latest jugg model just for testing.

float though? not fp16?

wouldnt that make the file much larger

ok. so this works well for you anyway? these settings?

I have more options than you though

[rustic garnet]

I would have to check code, but is it even working for any model? Usually, Loras leave out a lot of components from the unet (e.g. there are no bias layers in lora). If your model relies on that layers you might fail when extracting the lora

I would say for proper lora extraction you have to fine-tune the model on the same layers you later extract in the lora (usually just the attention layers)

so idea behind lora extraction is rather that you don't have to decide for a rank before training

[icy brook]

thanks for helping but I don't get the tech lingo. I am trying to extract a 128 dim lora from the latest Juggernaut model but just get bad output or no working file.

[rustic garnet]

I could imagine that it just does not work if the model was not initially made for lora extraction

like if you have a model that finetunes the unet and the text encoder and later you only extract the unet it very likely won't work.
Fine-tuning usually involves training all layers of the unet (and maybe text encoder) while loras usually only finetune a subset of the layers

it's not a technical limitation though, rather an implementation issue

[icy brook]

yes. I myself never train the text enc for my Aetherverse models. Jugg probably has trained the text enc though.

I suppose I'll just have to balance and merge between models

[rustic garnet]

text encoder should work. I just say there are several components in the full fine-tuning that are not extracted in the lora. If your model deviates a lot from base model it might break

[stone fossil]

A raccoon screaming at the PC screen, with the text balloon: "FUCK PYTHON!!"

https://civitai.com/images/11321886

https://civitai.com/models/424460/paradox-3-sd-xl-10

[ashen otter]

Aaa!

[safe stump]

This is cool

[ashen otter]

@glass forge 👀

[uncut steeple]

Using PixArt as a base and refining with unsampling with SDXL, really good stuff

[gloomy lark]

Can you paste a workflow doing what you're talking about? I'm doing regular ksampling with sdxl as a second stage but I'd like to see how your process differs.

[uncut steeple]

Sure, ive also uploaded it on civitai but here you go!

[gloomy lark]

Awesome thanks. What's your civit page? I'll add a follow

[uncut steeple]

My discord name, I dont like to advertise myself here

[gloomy lark]

Sure sounds good

[uncut steeple]

Unsampling is like sampler with flipped sigma, so it reverses the process until a certain step and reruns it but keeps the composition. Its really cool for refining stuff

Cause pixart itself is great for composition but details suck

[gloomy lark]

Clownshark has kept telling me how great it is, but I haven't had the mental bandwidth to try it out yet. Totally agree about pixart. My hope is that if sd3 is delayed enough, maybe someone big will do a fine tune of sigma

[uncut steeple]

Same, saw some people talking about fine tuning it but havent seen anything promising yet

[uncut steeple]

I think I talked to him as well about it

[gloomy lark]

On the pixart discord server I see people actively working on fine tuning and it's working, but i don't know how big their image sets are.

[uncut steeple]

Saw that too, but nothing to crazy from what ive looked at

Just as an example, pixart vs refined

[gloomy lark]

With the unsampling? Definitely a good output.

[uncut steeple]

Yep

I run it back 20 out of 40 steps and rerun the 20

[gloomy lark]

The whole face looks more normal as well

[uncut steeple]

You can also run it down to like 35 out of 40 steps, need to play around a little bit to get the idea

[uncut steeple]

Same here yeah, even tho the face still looks a little deformed

And the random hand appearing

Thats the issue with unsampling

[gloomy lark]

Sounds like what can happen with a denoise setting that's too high on the sdxl stage. I usually use 0.5 but I've lately been doing 0.4 so the random stuff doesn't happen

[uncut steeple]

Yep, its trial and error pretty much, but I like it over the basic img2img

[gloomy lark]

I find I'm able to do 0.5 when using style transfer with ip adapter. Because it also does a bit of composition transfer, it keeps things from getting out of line

[uncut steeple]

Maybe I should try a combination of style transfer and unsampling

[tawdry current]

Yeah that's what I was talking about the other night when I was showing how to use an unsampling workflow. You can mix other things into the ksampler half after it has been unsampled. But you just have to be careful about it

[uncut steeple]

Will try more tomorrow, but the first results were promising

[copper kraken]

yeah, i can't say enough how interesting unsampling is

it's basically a way to create non-random "noise"

to control its structure beyond shit like the distribution, color, scale, etc etc

[smoky shard]

Reference only half way through generation

[tawdry current]

Yep, it picks a noise seed and pattern essentially. If you don't unsample to 0, it will be partially solved. If you're using the same sampler type on the resampling half, there is no seed control(random will produce identical results). You can introduce variance by messing with the latents though like slerping latents in-between, but it gets messy sometimes and I think you have to enable the normalize flag on the unsampler since any random latent noise you might lerp it with will be normalized. I'm sure there's some way you could take the added noise and scale it to match the appropriate sigma level to match the unsampler's output though

[copper kraken]

"random will produce identical results" what do you mean by this?

are you referring to whatever seed is used with the flipped sigmas/unsampling?

[tawdry current]

Like if you have an unsampler plugged into a ksampler advanced. on the ksampler advanced, the seed value won't do anything

you have to perturb the second sampler in other ways

[copper kraken]

oh i always run without added noise

[tawdry current]

yeah you cant add noise on it

so you have to either find a way to inject noise between the unsampler and the ksampler, or you can just manipulate it through adding empty tokens to the second prompt

to randomize it

[copper kraken]

oh wait wtf it does have an effect without added noise, somehow i didnt notice that

prolly cuz i was using the same sampler when i tested that. interesting

[tawdry current]

if you have add noise enabled on the ksampler advanced, you will break things depending on how early you start the resample. like if you're doing 30 steps and the unsampler is set to 25(rewinding 5 steps), then 5 steps of added noise on the ksampler half isn't going to do jack because the sigma values are so tiny for the 25-30 range

[copper kraken]

btw, do you know wtf the noise is latent option is about?

[copper kraken]

definitely

[tawdry current]

noiseislatent is somewhat complicated

[rare axle]

Jo bosses quick question yall make those stuff in like the ‘normal’ stable diffusion?

When I generate stuff they look so clumpy and “bad”

Are it settings or like models that do the trick or am I missing something

[copper kraken]

any example use case? ive yet to find one

seed doesnt matter here

[rare axle]

what is this that u use if I may ask?

[tawdry current]

it normalizes them to latent space ranges i think:
if noise_is_latent: noise += latent_image.cpu()# * noise.std() noise.sub_(noise.mean()).div_(noise.std())

[copper kraken]

same with sde for unsampling

ahh

[tawdry current]

Yeah that's what i keep trying to say, any seed after the unsample will not affect it. The only way you can get variations seems to be through your secondary prompt. You can add in prompt noise with ,, ,,,,, type stuff

[copper kraken]

yeah its certain combos of samplers where the seed has an effect

[tawdry current]

but you will only get one seed out of it

with other samplers

[copper kraken]

seed changes it here

[tawdry current]

dont use ancestral with unsampling

only use convergent samplers

[copper kraken]

yeah i dont usually

just screwing around here

seed changes it here too

[tawdry current]

dpmpp_sde is ancestral

[copper kraken]

thats the sampling side

[tawdry current]

run it a second time and see if it changes

[copper kraken]

no seed control here

[tawdry current]

main point, to avoid any further confusion: don't use ancestral samplers. any other sampler you use might give you a different result, but usually only on the first try

[copper kraken]

the ancestral ones are fantastic with the iterative unsampling schedules

the determistic ones are pretty bad with that

[tawdry current]

again, it depends on the range your trying to rewind and resample

[copper kraken]

no seed control here... guessing the seed control is just if you have ancestral on the sampling side?

typing one handed with dvorak with a cat in my arm hence the incessant screen caps lol

[tawdry current]

well as long as the schedulers stay the same, it probably isn't a big deal if you use ancestral samplers or not. i'll have to play around and double check

but afaik, you can't switch out schedulers, since they will screw with the expected sigmas and whatnot

[copper kraken]

yeah for sure noticed that too

ancestral ive foaund are okay if you want to change the image and/or add detail without the full effect of regular denoising

[tawdry current]

oh i know why i said dont use ancestral, it's for the unsample side

[copper kraken]

generally i avoid unsampling ancustral

ahh yes

[tawdry current]

been a while since i thought about why i do things the way i do them lol

figured all this stuff out months ago

[copper kraken]

basically just denoising then

ahh nice

[tawdry current]

yeah, but you have to keep the same sigma schedules, you can use whatever other settings you like on the resample part

[copper kraken]

hey glad you brought this up, its stupid obv but i didnt think about the fact ancestral are affected by the seed without an initial denoising

yeah changing the shape of those is bad

ive found you can get away with diff total step counts on the unsample vs sample side

and skipping a step can be interesting

[tawdry current]

i like what it did with the ears lol

swapped it from rabbit to tiger

but omg thanks for rereminding me that i can just use ancestral samplers on the resample half, i was driving myself nuts trying to screw with latent noise blending, when it's this damn simple. i kind of remember running into the same issue months back and then having the eureka moment lol

(for getting random new variations)

[copper kraken]

thats awesome

100%|████████████████████████████████████████████████████████████████████████████████| 100/100 [00:10<00:00, 9.60it/s]
100%|████████████████████████████████████████████████████████████████████████████████| 100/100 [00:16<00:00, 6.24it/s]

wtf

i randomly got crazy speeds ive never seen on an unsampling run

9.60 it/s

cant reproduce it

wait i can

what the hell

workflow

whoa wait wtf cfg strength changes your it/s???

50% speed boost with cfg = 1.0000

6.2 it/s or so with cfg = 1.01 or 0.99

9.5-10 it/s with cfg = 1.00

seems to hold with every sampler

[tawdry current]

yes, at 1cfg, you have no negative prompt

so it's massively faster

that's why all the turbo/lightning models usually use 1cfg

[copper kraken]

ahhhhh

but the nef prompt is still used at 0, etc then

[tawdry current]

another example of it working pretty well

sec i'll tidy up the workflow a little and post an image so other people can try messing around

[copper kraken]

great result

[tawdry current]

kind of been annotating it for people anyways, since the workflow is super annoying to remember all the rules of lol

[copper kraken]

theres a lot of tiny deets for sure

[copper kraken]

one is the unsampler node can be crazy sensitive to the nef prompt

but not if you flip sigmas via custom

[tawdry current]

in the workflow, i'm generating the initial image with another model, but people can do whatever they want for that part or load an image

the only real limitation of unsampling/split sigmas type workflows is that you can't really extend the total number of steps. like if you are doing 30 steps and unsample to 25 steps, you're only going to resample 5 steps with the ksampler. and due to how sigmas work, you cant easily just magically extend that to 50 steps for that half

[copper kraken]

you can usually just scale

with that id do 180 steps rewin d to 150

but it is def different

cuz the steps arunt tiny

usually for the best ime

[tawdry current]

won't work because if you have some sigma curve like an exponential falloff or something, it gets stretched from 0 to N steps, like 30. now let's say you only unsample 5 steps of it. on the ksampler, you're wanting to stretch it to 50 steps, instead of to 30 where you'd only be resampling 5 steps. that sigma curve gets interpolated to the max steps, so to 30 and 50 in this case. the ksampler wants to start at step 25, well now the sigmas don't match correctly and it will likely result in there being too much noise added now

[copper kraken]

idk its actually worked pretty well for some stuff at least

[tawdry current]

yeah it probably will in some cases, but mathmatically, it doesn't

even if you scale the sigmas, to match what it's actually expecting before being extended, the rest of the sigma curve is still off

[copper kraken]

oh yeah with that case it wouldnt work, non-integer step schedule

i keep it to integer multipliers

usually i try to keep the number of steps rewound the same as the steps for the initial denoise

with that one youd need to stop at step 41.666666666666666666666666666667

[tawdry current]

like using this example curve here, from some random SD blog, the curve will always look like this no matter how many steps you add. that 20 could be 200 and it would still look the same, since it gets interpolated to length. but if you're using your own custom curves and whatnot, that's different, but experimental

[copper kraken]

yeah

[tawdry current]

so this is why you can't easily extend the resampler half of an unsampler workflow. However, you could always do an extra pass where you just denoise everything by .25-.35, since the scene composition is mostly good now, you just want crisper quality on it.

[copper kraken]

oh i blow up both parts equally... usually

[tawdry current]

and that's why i usually do it if i'm only getting to resample 5-10 steps out of 30 or something(microvariations basically)

[copper kraken]

yeah or use a sde on the sampling side... res_momentumized can be really really good for that with a custom sched

30steps rewind to 25 for both

first, resample 25->30

second, resample 150->180

three seeds with res on the sample end

ffs cant believe i didnt realize those seeds matter 🤣

very glad you mentioned that

[tawdry current]

Yeah see you're still maintaining the same percentage spot through the noise schedule. There are more steps, yes, but it's usually going to produce very similar results with a hair more accuracy. What I was talking about would be like if you wanted to go from 1-(25/30)=16.7% to something like 1-(25/50)=50% if you wanted it more in terms of a relative denoising percentage

for more potential changes to the image, but that's why i was saying it's easier to just add in an extra stage where you then resample the whole image at .25-40 denoise

[tawdry current]

out of all the endless things that can be made with stable diffusion, why are you making some 12 year old looking child?

[fierce heron]

cause she's infinetely more pretty than anything you've ever made 😂

[tawdry current]

And how old are you?

[glad grove]

the stain he left at the sofa is pretty too

[fierce heron]

64

[minor pecan]

[fierce heron]

👵

[copper kraken]

ohhhh

i see what you mean now, yeah

thought you meant something else

[tawdry current]

A 64 year old sitting around making hyper-realistic pictures of 12 year olds... Yeah, I don't think anyone really needs to actually spell this out for you

[glad grove]

you should spell it out for him since you are the expert here

[tawdry current]

Why don't you take the conversation to a more public forum and not some tiny echo-chamber and see what the macro-consensus is on the topic...

[glad grove]

what macro consensus the one when you end up looking like a clown?

[fierce heron]

i dont think the macro-consensus is that making art of youthful people is wrong

otherwise i think you should call disney and tell them they are all pedos

[glad grove]

yea i dont see any 12 yo here only 12 i see here is that guy iq

[gloomy lark]

That's just a photo of some barely pubescent girl. That's not art.

[tawdry current]

exactly

[fierce heron]

AI art isnt art. got it

[minor pecan]

https://tenor.com/view/steven-seagal-seagal-knife-fight-gif-18480228

[gloomy lark]

That's some serious deflection there.

[glad grove]

almost a giant reach as trying to claim some picture age

[glad grove]

https://tenor.com/view/seagal-steven-seagal-carrot-eat-gif-13722967

[tawdry current]

Oh wait, let me guess, they are 1000 years old, they just look like a 12 year old... right...

[glad grove]

i mean it works for you