💬｜general-chat

[mortal delta]

that sounds cool the work you did, but its a real bummer its overdone now.

[karmic cedar]

Profoundly accurate 😔

[shell tendon]

it's not so much that it's overdone as it's that ppl are having massive irrational panic attacks and getting very aggressive with some content creators

[mortal delta]

isnt that just the internet in a nutshell? practically?

[shell tendon]

yeah, and ppl in general in real life too, which is where half teh problem is

bet game devs etc are getting hammered with death threats every time they use generative AI to create something, in part because of the effect it has on jobs, when we've already been shipping jobs out of the country anyway etc...

[mortal delta]

oh i saw a post on reddit of someone who made a barebones point n click game with gamemaker and stable diffusion, and i bet soon enough people are going to panic about it or something weird like that.

[forest trout]

Some of us still remember the panic attacks that people had when Photoshop was created. Or the panic attacks photographers had when digital photography became a thing.

[shell tendon]

yeahhhh.

i wasn't old enough to have much memory of it, but i have family who have done a lot of professional work in photography/film/etc... one was one of the first to make truly convincing photoshops and caused a real shitstorm in their field

stupid how much blowback there was

[forest trout]

People didn't like the idea that you didn't have to develop film. There was plenty of pushback because "digital photography will never be as crisp and clean as analog".

[shell tendon]

yep that's what i was told they were blasted with

won't have the same rich colors etc

reminds me of the same crap with vinyl

[forest trout]

Also "If you use samples in your beats, you aren't a real musician. You have technicians making noise, no one here is a musician, because no one knows how to play the guitar!"

[shell tendon]

can't count how many times i had ppl tell me that they could tell the difference between a really well encoded mp3 and a .wav on their headphones... and that vinyl is a more accurate representation of the original sound than digital

haha yesss

like the ppl who used to get really upset when someone would program in a midi loop for a synthesizer using a mouse

[stable comet]

Hello all. I discovered the world of AI generators a few days ago learning about it. I chose to got with the stable diffusion route even though i don't have a GPU. I know it will be painstakingly slow without GPU and maybe I won't get far but wanted to delve into this world that I am VERY interested in. If I am able to compute with this stuff I would invest the $ to get the proper hardware. With that all being said I am in need of help for a new beginner. I can't seem to get the stable-diffusion-webui.git to install the necessary files as I get a return error of https: is not recognized as an internal or external command.. I am the proper folder \Automatic1111> and copied that line in and this is where I am stuck. Thank you for your kindness and understanding of my struggle with it. I was following a YT video verbatim and there doesn't seem to be any rhyme or reason for my not being able to move forward. Thank You.

[shell tendon]

you'll want to take your question over to #🤝｜tech-support , there's ppl in there who are very helpful 🙂

[stable comet]

Thank you very much! I'm new to discord too. Appreciate you.

[past charm]

[sweet chasm]

[sage reef]

i wish they would release stable audio 2 as a local thing too (aka the weights and inference code), cause I would love to use it, i know they said they are working on some version that will be free, but im not sure if that version will be on the same "level" or quality as stable audio 2

[pearl ocean]

Me when SD3 releases

[vale steeple]

I know it is trained on different dataset

instead of Audiosparx

but what dataset? I dont have a clue

[opal hedge]

I really wish they'd release progress updates about sd3

[winter sedge]

anyone having htis error with animatediff? video_length = mm_animatediff.ad_params.batch_size
AttributeError: 'NoneType' object has no attribute 'batch_size'

video_length = mm_animatediff.ad_params.batch_size
AttributeError: 'NoneType' object has no attribute 'batch_size'

animatediff error on stable webui and forge webui

[shell tendon]

they don't do that with anything

[heavy lark]

based on tweets and reddit threads, some who had it originally, no longer do. some who didn't have it originally, now do. so things are moving.

[karmic cedar]

sd3 when?? ☺️

zero point free energy when???

[stuck sun]

why would they revoke access like that

was it the guys on civitAI who generate nothing but various fluids seeping out of engorged anuses trying to get SD3 to do nothing but that?

and Stability were like guys I know you are red teaming it here but can you try to generate anything at all other than fluids seeping out of engorged anuses

and they were like "no. We cannot commit to that."

[trim magnet]

why do u know they generate that and why do you care if they do

[stuck sun]

have you been to the civitAI website sir

[trim magnet]

plenty of times

[stuck sun]

ok, my follow up question is do you have functioning eyes

[trim magnet]

do you?

[stuck sun]

yes, this is the reason why i know they generate that. Because i have been to the website and have functioning eyes.

[trim magnet]

then
a.use the nsfw filer or
b.dont go there

[stuck sun]

why do you care that i care

[trim magnet]

why do you care that i told i you that,you can try to cope with it

[stuck sun]

sir please, this is boring.

[trim magnet]

the door is open

🚪

[heavy lark]

probably limited resources. and to be straight forward, at least the stuff that gets posted on twitter, is usually the most boring benign stuff I've ever seen. 99% of it doesn't in any way exercise what SD3 can do.

[stuck sun]

yeah every image i've seen posted had that ugly blasted contrast/oversaturated/ultrasmooth skin look that plagued SDXL too, very ugly aesthetic. Text handling was better i guess, but who knows how much they are cherry picking.

[heavy lark]

A couple pics were generated where I asked for them, and those were impressive. but pretty much everything else was a girl standing there. a small figurine. some water color jellyfish. argh.

meanwhile with dalle "show me a picture of a universe made entirely of cheetos, where matter itself has cheeto atoms made up of cheeto subatomic particles of cheetos" and it does it. "SD3 research paper: jUst As geWd as Dall3!"

[stuck sun]

And MJ is miles ahead aesthetically

anyway I remain hopeful!

anyway I remain delusional!

[heavy lark]

well, i'd say with ipadapter, all that MJ styling is easily attainable in sd3 now

[karmic cedar]

https://www.reddit.com/r/StableDiffusion/s/dJpuPy7qOV

Regarding aesthetic latitude

[shell tendon]

yeah, it really is frustrating how many dumb prompts have been thrown at SD3

i'd like to see "woman standing in kitchen holding a bloody knife behind her back"

[heavy lark]

I tried every which way with sdxl, all models, i made it a toy instead of a knife. closest I could get is a view from behind.

[rain sphinx]

[pearl ocean]

Should try Epicrealism

[heavy lark]

lol i pulled it up on civit and it's just wall to wall porn

[urban loom]

Hi Everyone. Not sure if this is the right channel for this. I'm looking for a stable diffusion / MidJourney professional who can assist with a project on digitally altering images of socks. I have PNG images and 3D files of the socks. The goal is to take these images, keep the socks unchanged, and completely transform the model and background to a design of my choice. I would also love to learn this process. If anyone is skilled in these techniques and is open to collaboration and teaching, please DM me. I've attached an example of the final result we want. Thanks!

https://media.discordapp.net/attachments/1002602742667280404/1228940618978295878/image.png?ex=662ddf24&is=661b6a24&hm=b55f4e254078053c2be7e022c09aa3567cae8184324743a23acd79258bf9a60c&=&format=webp&quality=lossless&width=1440&height=486

[stuck sun]

honestly sounds like it would just be easier to do in photoshop

[urban loom]

I would be happy to know how 🙂

[pearl ocean]

MS Paint

[stuck sun]

object selection tool makes cytting out objects easy

transform/warp lets you reshape an object

use a few feathered masks and curves adjustments

and its about 2 mins work

https://www.dropbox.com/scl/fi/du3xq5kbrsz1q7hxckcca/socks.jpg?rlkey=uve7igw5q7zy1e14d17sjk3w3&dl=0

[urban loom]

But first I need to create the new image. I can't copy other people images and backgrounds 🙂

[stuck sun]

but you can use this to put it into any image you want

people wanna use AI for everything and it aint always the simplest solution

i mean otherwise you take a bunch of images you want and train a lora with that sock design and generate whatever images you want with these socks in 'em, but its a lot more work and trial and error, etc etc

or if you have 3d files of the socks and know your way around 3d you can generate the images you want in SD or MJ and render/comp the socks in if they need to have very specific lighting

etc

you said you have 3d models right?

someone made a socks adetailer model

https://civitai.com/models/311874/socks-segmentation-adetailer-sockssegsyolov8

[vale steeple]

https://www.reddit.com/r/StableDiffusion/comments/1c2je28/i_got_access_to_sd3_on_stable_assistant_platform/

This thread got a fair shares of difficult prompt

to be honest

[stuck sun]

interesting, there's a pissy exchange between a guy thats critical of the model and a guy defending it and i dont believe the guy defending it for one second

the "its good if you prompt it good" guy

[crude notch]

huh i wonder who that could be

[loud solar]

My expectations aren't that high any longer ...

[crude notch]

(the model is good if it is prompted good)

[crude notch]

like 2-4 weeks

[loud solar]

The last release wasn't really good. Swarm intelligence made it interesting

[crude notch]

what do you mean by "swarm imteligence"?

[loud solar]

Stable Diffusion release didn't reach expectations. Community models made it usable ...

[crude notch]

eh, its almost always like that, sd makes a great base, people finetune to specification (weeb, not weeb, photos, whatever)

[loud solar]

Didn't they show SDXL can do text?

[stuck sun]

I think my least favorite kind of guy is the "why wld u want to do that?" guy that infests literally every forum for every piece of sofware or hardware and naysays anyone who asks for any kind of improvement ever, these people are put on this earth by satan to drain everyones will to live. The model is good if it is prompted good guy is that kind of guy.

[crude notch]

sd3 is good if prompted good, if you prompt sd3 like dall3, it will not be as good

tag based prompting is dead (for launch sd3)

[stuck sun]

"of course its good for multi characters, anatomy, and complex poses, it works great for me, here's an image of a person standing there is a really boring static pose to prove its good at that."

[crude notch]

i never said that, though will test if you have your doubts

[stuck sun]

?

i'm talking about the guy in the reddit thread

[crude notch]

ParseeMizuhashiTH11?

thats me lol

[stuck sun]

lol

[loud solar]

A lot of pictures they show should be able with an older version, too. Maybe with additional tool ...

[warm relic]

im reading the threads. if the research team departed from stability, who is touching the model currently? so everything testers are playing with is no longer the original model produced by the research team but rather something still being messed with by, who?

[sterile gust]

No hope for SD3 and i doubt the community can fix it.

[stuck sun]

here's a prompt SDXL totally fails at

analog photo of 3 women doing a backflip in mid-air, dynamic pose, motion blur, vintage, faded film, film grain, wearing tracksuit, soaking wet, heavy rain, sunset, flare, light leaks,

helloworld:
https://www.dropbox.com/scl/fi/ot7qcqf9chrieqcq58giy/grid-0334.png?rlkey=7q3nn69qsaehzod7augv39kbx&dl=0

Juggernaut:
https://www.dropbox.com/scl/fi/66li56s63we5et88gxmb3/grid-0335.png?rlkey=peqjo0a1kavw356yxrjacqzm5&dl=0

SDXL Base:
https://www.dropbox.com/scl/fi/8ug17bmwmbtbslbo258rn/grid-0336.png?rlkey=8ze1w8mc2kxam2z7cv4d4fyhd&dl=0

in every case it turns into a mutant freakshow because XL doesnt understand human anatomy in extreme or complex/dynamic poses

lol

https://www.dropbox.com/scl/fi/njgixwmhczimdul7aztcu/03415-3531789564-analog-photo-of-3-women-doing-a-backflip-in-mid-air-dynamic-pose-motion-blur-vintage-faded-film-film-grain-wearing-tracksu.png?rlkey=xg73fy6mm5iohr9ocq4n9rf0z&dl=0

[shell tendon]

take out everything but "photo of 3 women doing a backflip in mid-air"

the other shit muddies the question of whether it can address that one core concept or whether it's just losing focus

[stuck sun]

why, i want it to be sunset and raining and look like a vintage photo, and the sun to be flaring into the lens

[shell tendon]

the big issue is SD isn't trained on upside down ppl

yes, you do, but if you run into a problem, best troubleshooting strategy is to reduce the problem down to what you think the key issue is

reduce the variables to a minimum so you can come up with firm answers to questions

[stuck sun]

sure but what i want is for it to actually look good too

[shell tendon]

right now it's not clear if another one of those tokens is causing some issue

[stuck sun]

i dont care if it can do a backflip in bad lighting

[shell tendon]

i sure as hell do

cuz that i know is where the issue is

if you don't figure out what the problem is you also won't know if you're just slamming your head into the wall

here's the issue: SD isn't trained on people being upside down

[loud solar]

It can work with sports ...

[shell tendon]

there may be very specific cases, but try denoising a face after rotating it 180 degrees

in general it will do crazy shit

[stuck sun]

congrats we now have analog photo of 3 women doing a backflip in mid-air with bad lighting instead. https://www.dropbox.com/scl/fi/2s8hwsvs708q4ym9frtff/grid-0339.png?rlkey=chwhnd3opu3ezuuklxsvlixdv&dl=0

[loud solar]

lol

[shell tendon]

and that's important you do that

now you know it's not the lighting, or another token that's causing the issue

[stuck sun]

I knew that already because i added the tokens one by one.

[shell tendon]

there's many cases when adding random tokens will cause another unrelated part of the image to get fucked up

yup

so, there's the problem: upside down

the faces are f'd

now if you know that, then you can generate them right side up via denoising the f'd up version there

then rotate it back

[stuck sun]

yeah but also the other problem isn't just a person upside down, there's a whole arc to it with multiple stages and recogniseable poses and it doesnt understand any of it

and it doesnt understand motion blur or what arc any of the body is traveling on

[shell tendon]

yeah if you want to do the entire thing frame by frame that's beyond me

my knowledge there ends with the static image approach

[stuck sun]

i'm talking about static images, i'm talking about capturing dynamic poses

[rigid hull]

Is anyone who knows how to install stable diffusion in low end pc

[stuck sun]

like here's a use case, people wanna use this stuff to do storyboards right? its a common thing in tv commercials now, people do boards in midjourney

so you need to be able to do dymamic poses for that

cinematic photo of a Boxer being punched really hard in the face in back-alley brawl,

doesnt know what a punch is

[shell tendon]

yeah, that's def openpose and pray

yeaah

it's really bad at that

gotta do it manually

[stuck sun]

yeah but thats the problem! Because then there's a whole chain of issues downstream from that, like having characters interact in complex ways, and not just porno ones

[lucid frigate]

i am trying to install forgeUI for automatic1111, but while i was installing it says "CUDA Stream Activated: False" in the terminal. how do i activate it? i already installed it but i dont know how to activate it

[shell tendon]

yep, that is the problem, lol

[crude notch]

it works (just the photo of 3 women doing a backflip in mid-air part)

i am, not, prompting the rest (out of concern for my access)

[stuck sun]

the rest meaning the faded film stuff or the punch?

because the punch = violence and there's rules about that?

[stuck sun]

hey btw sorry if i was kinda a dick, a couple of posts on reddit rubbed me up a bit the wrong way but I appreciate you taking the time out and not being rude back.

[loud solar]

My brain just filters everything that shouldn't be there ^^

Uhh... that sounds kinda ignorant ^^

[stuck sun]

its so funny, MJ will do the prompt "cinematic photo of a Boxer being punched really hard in the face in back-alley brawl, breaking his cheekbone" -- but it works really hard to avoid showing the point of contact, it does all the big dynamic motions and pained expressions and sweat flying through the air but either the blows are like a boxing glove being nudged up against the face or a guy looking like he's punching himself, It knows what a punch is but it will NOT do the heavy contact, because its so censored.

https://www.dropbox.com/scl/fi/3x4jsvxpy38c57scwo3v6/MJboxing.PNG?rlkey=iz9oyv7bq5hr4aal620bwvubl&dl=0

it did one weak blow out of about 50 gens
https://www.dropbox.com/scl/fi/c26kcco9c09my8agvg4qh/0_0.webp?rlkey=t6fx1ie57srmtk3wpnfaw6fxp&dl=0

and the same prompt in SDXL is just two sweaty guys posing next to each other looking tussled, because it doesnt know what a punch is
https://www.dropbox.com/scl/fi/3tt4rpomc2nojc2pvubf5/grid-0342.png?rlkey=c8bqy10l1svrvfwisen7i6seg&dl=0

[loud solar]

Thank god my brain censors itself ... so I don't get problems like that ^^

[stuck sun]

problems like which?

[loud solar]

Bad punches ...

My biggest problem is getting not so male nude males ... but that's not really a big issue

[stuck sun]

yeah SD is definitely weighted toward generating ladies

[warm bane]

Hi guys, what do you think in your opinion is the best upscaler?

Something more natural looking like a digital drawing

[stuck sun]

Supir

[warm bane]

I really like Anime 6B or whatever but im looking for something even more vibrant

[warm bane]

Supir Upscale? (ComfyUI) does that mean I need ComfyUI for it??

[stuck sun]

uhh I have a standalone installer for it

not sure about comfy

[loud solar]

Didn't make it work in comfy but not really that used to it ...

working with ultramix_balanced at the moment

[stuck sun]

Dr. Furkan Gözükara made a standalone supir upscaler 1 click installer, i had to join his patreon to get it tho, not sure if he's released it publically since or there's another way to use it in comfy or whatever, but anyway, best upscaler i've ever used

[warm bane]

interesting, im looking at it and it seems amazing but im mostly looking for one that looks like a natural digital drawing

this one seems to be focused on realism ?

[loud solar]

It's for his patreons only ...

[stuck sun]

yeah he tested a few illustrations in the video

[warm bane]

can u link me vid? 😄 I'm curious to see it

[stuck sun]

possibly not anime style like you want but i guess i could test something for you now

[warm bane]

Also, do you guys have any recommendation for a good model for Fantasy Digital art

[stuck sun]

https://www.youtube.com/watch?v=PqREA6-bC3w

[warm bane]

that looks natural + vibrant colors

[stuck sun]

gotit, yeah i havent touched it in about a month so had no idea if there were other versions released or whatever since then

[loud solar]

SUPIR uses an XL Model ... you could try an anime one ,,,

[stuck sun]

oh yeah yes thats true

good point

i've only used it for realistic stuff

[warm bane]

yeah im not looking for realism

[stuck sun]

yeah but not exclusively for realism

he tests a 2d image if you skip fwd here:
https://youtu.be/PqREA6-bC3w?t=1564

[warm bane]

yeah that one looks great

[stuck sun]

the value of supir is it preserves the integrity of lower res images incredibly well, subtle facial expressions etc, every other upscaler i've tested changes way too much stuff

so if thats important its amazing

[warm bane]

yeah thats true, the colors look good but it doesnt keep as much detail

[stuck sun]

if you want it to add detail or fix things maybe its not what you want

[warm bane]

anime 6b even though it looks good kinda looks like some heavy brush

[stuck sun]

althought you can still do that if you dial in settings and prompt etc

[warm bane]

what about a model for fantasy digital art ?

[stuck sun]

because there's denoise strength type controls too

[warm bane]

https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/2cd83a11-2952-47a4-3db1-d493f115d700/original=true/147531.jpeg

Im looking for something like this

[stuck sun]

sorry cant help you there, not been my focus

[warm bane]

thanks tho this upscaler looks good

seems like it preserves stuff and just upgrades the image quality

[stuck sun]

yep

kinda needs a beefy card tho

12GB vram minumum

[warm bane]

ye i got 12 gb exactly

[stuck sun]

at least it was last i used it

they may have updated it since (when it first came out needed 32)

[warm bane]

oh one thing I'm curious about when you see very good images generated on CivitAI

Why do they not include the upscaler too?
Do they really generate an image that good without using upscaler?? 😄

I thought most models can't really generate properly in such big resolutions

[warm junco]

They also contain the upscaler version. But its only visible in the metadata. You have to click
"Copy prompt", then paste it into your prompt field und press the Blue/white arrow to let it automatically set everything the same

Do you need good upscalers for Anime?

[warm bane]

not anime, just looking for some model and upscaler that does more natural looking HQ digital art

[warm bane]

oh really, interesting, ill have to try that too

is that Copy Generation data basically?

[warm junco]

NMKD Siax 200K is a good upscaler

[warm bane]

[warm bane]

thanks, and where do i paste the generation data exactly ?

[warm junco]

In the prompt field

[rugged vale]

what kind or regularization images should be used for a style?

as in lora training dreambooth method

[storm shard]

Or do it yourself, it seems to work (I haven't tried it)
Local Supir
https://www.reddit.com/r/StableDiffusion/comments/1b37h5z/supir_super_resolution_tutorial_to_run_it_locally/?rdt=36399

[hexed pilot]

Been hearing lots of positive stuff about SD Forge. I'm pretty new to all this, is Forge like a stand alone version of SD, or do I need Forge+sdxl for example?

[warm junco]

Forge is a webui like automatic1111 or comfyui.
It has some advantages for GPUs that haven't much Vram.
All the webuis can work with 1.5 and sdxl models

[wet dawn]

hi

[hexed pilot]

I see, so if I want to install and use Forge, I need the baseline sdxl and Forge, if I understand everything correctly?

[hexed pilot]

Hey, welcome

[primal trench]

Any sdxl model should work. Even trained/finetuned ones

[knotty ferry]

Hi everyone, no support received in the past few days.
StabilityAI has generated some inappropriate images that I would like to delete from my profile, or close the profile if this is not possible. Would it be possible to receive support? Do you know how to do it? I sent several emails but never received any replies thank you all

[civic schooner]

What do you mean by "your profile"?

[knotty ferry]

I made a profile here to generate images I also pay a subscription https://stability.ai/

[fervent thunder]

anyone knows how to use TTPLanet_SDXL_Controlnet_Tile_Realistic with a1111 ?

[hexed pilot]

Holy crap this is giving me a headache Trying to categorize all of my wildcards. Anyone wanna share how you've organized it?

[knotty ferry]

Can anyone help me get support from stabilityAI? Thank you all 🫶

[hexed pilot]

Would if i could brother

[marsh hound]

Where are the robots? HOW CAN I GENERATE PHOTOS NOW?

[amber bloom]

You must use the power of your mind

[marsh hound]

PLease guid me, I haven't used this server in a while

Before it used to be 10 bots

And you could use dream to make the photos

[amber bloom]

The bots are gone for now, maybe forever. Try using https://leonardo.ai/ if you want to generate photos with stable diffusion

[trail lion]

or install stable diffusion on your pc

doubt the bot ever comes back frankly

why in the world they were letting users generate images for free on their resources when they couldnt pay their cloud providers

[compact swan]

just installed stable diffusion on my pc

auto1111 or whatever it's called

it's pretty good but batch size 4 and 10+ sampling steps takes absolutely forever and uses so much VRAM

I think I'll need something better than a 3060 Ti

[oak oracle]

😭

[karmic cedar]

https://youtu.be/rZC65Q3F_Mk

nuts

[storm shard]

just don't use batch size 4, use 1

[oak oracle]

wrong chat moment

[compact swan]

yeah i found out that batch size 1 and batch count 4 is better

because it makes all the images serially instead of in parallel

[lucid trail]

Does anyone know the best way to apply a style LoRA onto a photorealistic image? I've tried multi ControlNet, T2I, different SD base checkpoints, different style LoRAs results are always uncanny and doesn't look great

[grizzled palm]

Someone sharing SD3 testing on YT. I assume there have been others but this was the first that I came across? Also tried searching, but I can't find anything. So here you go. A link to a Stable Diffusion 3 testing video.
https://www.youtube.com/watch?v=mQSKoAEaIJA

[crude notch]

no worries, i never intended to be mean to anyone else

[charred mesa]

@crude notch what do you have to say about amazing painter's responses

[crude notch]

they assume that sai wouldnt do something that can improve training

also the feedback on "are you using 77tkn t5" was kind of mixed

(they arent)

so i dont blame them for using that as an argument

[charred mesa]

hmmm

its weird cause in the paper there was like a super long prompt that didn't have distortions

which is around 124 tokens

[crude notch]

yep, thats the t5 pulling its weight

[charred mesa]

I mean 77+77>124

idk how they work together

[crude notch]

basically:
are you going to use text?: use all/t5
are you going to have a long prompt? use all/t5
else? use both clips

[crude notch]

they each get same prompt

the clip, the other clip, and the t5-xxl

[charred mesa]

no I mean the two clip models

whaat

but when I enter a prompt too long it errors out

when I add it to only one of the clip models

inside comfyui

[crude notch]

oh yeah, thats the 77tkn limit

clip can only handle 77tkns

[charred mesa]

but then what do these GUIs do when it comes to accepting longer prompts when using longer prompts

[crude notch]

but if you just condCombine(prompt[:77], prompt[77:) you bypass it

[charred mesa]

just automatically truncating or somehow concatting

[crude notch]

concat embeds

[charred mesa]

oh

so that's why

[crude notch]

the one flaw (which was posted in the comment thread) is that they dont share attn

[hoary lark]

is there any really good in depth guides for making ai generated images of a person you train into it with a couple photos? im trying to find one because im trying to train myself but its not working.

[crude notch]

so if you say "cirno [blah blah blah blah to fill 77tkns] | she [...]", the model will not know what she is

[charred mesa]

look up Lora tutorials, those are light on VRAM and work with few photos iirc

[crude notch]

lora? theres a bunch on civitai, though for hyperparams if you just use the default settings you should be fine

[hoary lark]

its just everytime i try it never gets my face right

[charred mesa]

I wonder if there's gonna be a way in comfyui to make it only use T5's embedding/conditioning

🤔

[crude notch]

yep!

[crude notch]

ah, faces, install adetailer

[charred mesa]

would that be messed up

[crude notch]

no, not that messed up

[charred mesa]

hmmmm

I wonder what the results would look like

welp, gotta wait those 2-4 weeks to find out!

[crude notch]

different.

[charred mesa]

well yeah hehe

[crude notch]

i think the paper has something about that

[thick cliff]

very good video, and the results are quite good. I hope to still be alive when they release the version

[charred mesa]

Also what's with the abundance of prompts without styles in them

like most of them don't start with "photo of" and etc

its so vague

[crude notch]

people dont know how to prompt

[charred mesa]

not just that, the paper itself sometimes does this

A car made out of vegetables.
a massive alien space ship that is shaped like a pretzel.

[crude notch]

huh

oh yeah they also have some toggleable styles so maybe that?

[charred mesa]

yeah they might not have used raw prompts

[pine shale]

how do i download for amd gpu?

[foggy halo]

SD3 RELEASE WHEN?? 😭😭😭

really interesting

[coral drum]

Hello

[charred mesa]

sometime in the future

[jade mason]

SD3 when?

[foggy halo]

same question 😔

[trail lion]

there is something satisfying about a lack of a defined timetable

[foggy halo]

i saw this video and there seems to be a separate server only for sd3

if someone had the link of the SD3 Launchpad server and would share it that would be great

[low moon]

Censorship! Waifu trainers get ready to do overtime.

[pearl ocean]

GTA6

[shell tendon]

I don't think there's really a shortage of art on this server

i alone have generated over 100k images since dec 7 2023

[rotund bobcat]

Hi there, does anyone have any idea which AI tool was used to create the filter effect in this video? https://www.youtube.com/watch?v=vxRY_9bv8sc

[pearl ocean]

thats a good start

[trim magnet]

quality over quantity

[pearl ocean]

https://tenor.com/view/why-not-both-why-not-take-both-gif-11478682

¯_(ツ)_/¯

[shell tendon]

yup

[trim magnet]

or in this case none

[ionic shuttle]

hi

[karmic cedar]

what’s your blood type

mine is pina colada

[trim magnet]

mine is obama (i bleed blue)

[olive sparrow]

hi

[keen turret]

https://tenor.com/bEz63.gif

[shell tendon]

#🏞｜general-with-images message you can have quantity and quality if you keep a 4090 humming around the clock

[nova zodiac]

I could sell my car twice and still not have enough $$$ for a 4090 lol

[trim magnet]

every artist oversells their own art

[karmic cedar]

all art oversells its artist

[trim magnet]

yea the banana one sold for a lot

[karmic cedar]

and the pee one

with the urine

and the piss

[trim magnet]

also the guy who destroyed the painting with a shredder

[karmic cedar]

that’s just considered time art

j/k that’s just being an artist

[pearl ocean]

the price went higher

[shell tendon]

you have me curious wtf you're driving

i could replace mine, but then i realize i could also buy 15 5090s when that comes out with the same amount of money

[nova zodiac]

98 nissan primera (infiniti g20 equiv) with 170k miles (272k km) on the clock

[shell tendon]

whoa, impressive

worst i've seen was probably my dads car

pushed an 89 accord to nearly 300k miles

ski rack flew off on the freeway and it was all downhill from there

[pearl ocean]

what would you do, if you have tge power of the M3 Pro?

[shell tendon]

by the time it finally died died, the drivers side window didn't roll down, neither back door opened, the A/C was dead, radio was dead, and you had to put a golf tee in a hole in the stick shift to change gears

[nova zodiac]

I've seen a couple of 1990 toyota corolla wagons hit north of 400k miles, have also seen a 2005 mazda 6 wagon hit 300k miles as well (work vehicles, not personal ones 😛 )

[shell tendon]

yikes

[nova zodiac]

yeah those Japanese cars are built solid!

[shell tendon]

friend of mine who's a doctor and has plentyyyy of money still drives this complete pile of shit truck he assembled from three different trucks he salvaged at a junkyard

doesn't even have a key for either of the doors, both of which are different

a few years ago i borrowed it for moving some stuff around town (it's stick btw) and the fn stick came off in my hand while i was rounding a turn

nearly took out a limo as a bride was about to climb into it cuz i was so stunned, my brain just start smoking pretty much when that happened

[fallow fog]

Hello Everyone

I made a high quality text-to-image dataset based on midjourney images providing two level of prompting (long prompt and short prompt for the same image)

https://huggingface.co/datasets/MohamedRashad/midjourney-detailed-prompts

I hope it be useful for someone here

[karmic cedar]

more ella-style attention guidance stuff coming out https://github.com/KU-CVLAB/Perturbed-Attention-Guidance

[fair pewter]

hello

using comfyui but having an issue atm

anyone down to help?

[sterile gust]

Hope someone has the card for that.

[nova zodiac]

there are a1111 extensions and comfy nodes out for this already 🙂

[boreal dew]

would anyone know why civitai seems to be using so much of my computer's resources? As soon as I open a tab, my laptop goes into overdrive with the fans

whenever i click off to another tab, the fans relax.

[shell tendon]

idk, just watch it turn out they've found a way to sell your cpu cycles

i've noticed the same thing

[boreal dew]

I can barely use it without my laptop getting way hotter than it should.

And I'm just trying to find outfit lora's...

Having it open and in the background doesn't do anything, but as soon as I click on it, boom. fans are mad.

[lapis sage]

morning

meow

oh the images are gone here, stay tuned in #🏞｜general-with-images

[boreal dew]

ahhh i thought i figured out a work around by just making the window smaller. damn.

it kinda works.

[nova zodiac]

poorly optimized website - best way is to not have it load images, but that kind of defeats the purpose lol - you can disable hardware acceleration and that does help apparently, but you are by no means the only person complaining about this

[fair pewter]

i mean

i had to read this 3 times to understand

it depends on the laptop

if your laptop is old then try to get it clean from the inside

because it shouldnt go rocket mode just for opening a website

and if it still does it consider getting a newer one or change the ram slot

since ram degrades overtime

[ionic oriole]

gm

[fair pewter]

morning

[pearl ocean]

Evening

[strange heath]

1324

[hasty arch]

gm

[full sandal]

hello

[heavy valve]

https://www.tiktok.com/@edwinskeletrix/video/7355974950945164586 how do people make videos like this? can i do it locally?

[hushed karma]

gm

[dawn warren]

gm

[sand tide]

gm

[zenith forge]

gm

[pearl ocean]

Evening

[nova zodiac]

I believe it's pretty easy to sit in front of a phone and talk like a derp..

[nova zodiac]

Wassup!

[pearl ocean]

Yo

[sudden ruin]

Average tiktok moment

[nova zodiac]

Really don't know how people get addicted to watching that shit

Either that, or I'm getting old AF

either way, the brain rot is strong!

[sudden ruin]

Same
Ive seen my gf spent half her day in front of it, its somewhat depressing

And I have zero understanding for it

[gaunt arrow]

hello

[pearl ocean]

Same, I’m not into TikTok and stuff

[proud spindle]

hello

[nimble violet]

gm

[flat blade]

gm

[fervent thunder]

Thats the point. You may not realize that there are levels to brain rotting and you and I are no exception, just due to the fact of being here

[nova zodiac]

Agreed, we just choose a different poison 😛

[fervent thunder]

Question is, why do we need poisons? Is 9-5 that miserable? (Work or school) 😭

[nova zodiac]

You mean the dystopia that is late stage capitalism.. nah no idea mate 😉

[fair pewter]

yep you can

[low moon]

So looks like SD3 wont be as good as Ideogram with words. But who cares sicne you cna poull the levers 100s of times on your own machine until it gets it rgiht

XD

[nova zodiac]

Someone will make an equivalent of Harrlogos for sd3 and itll be fine

[heavy valve]

using what

[low moon]

Only two classes left the have nots and the have debts.

[fervent thunder]

Does anyone how to generate an image given a small part of it? For instance given, flower with no background, generate someone holding it?

Outpainting of sorts but with custom shape

[low moon]

rich people dotn exist they cant access 99% of their wealth if they have any, governments after them or "invested" in "stuff" etc, middle class gone logn tiem ago, anyone who appears to have soemthign is in debt for decades to pay for it

What a time to be alive!

Two appers down theline maybe I can escape to Mars.

On ym Muskmobile.

[fervent thunder]

Yes its that 0.01% and the rest. Dont worry tho, soon ai will replace u and u dont have to work anymore.

[nova zodiac]

There was only ever two classes, the capital/resource owners and the workers/slaves

[nova zodiac]

Hold on to your papers!

[low moon]

I welcome our robot overlords. they can't possibly screw up worse than us.

[fervent thunder]

Unless the elites control them

[low moon]

Mainly Zuckerberg.

But he is alreadya robot, more machine than man

[fervent thunder]

I loved the quote of some guy in wef, that said we wont need the vast majority of people anymore, even as slaves

[upper geyser]

hello

[low moon]

The Georgia stones specify they only need 500 million of us to wipe their asses

so essentially the top epople advocate a genocide of 7.5 billion people

while they sit in their pools udnerground and wait for the radiation to disperse

Even top baddies of history can;t coem close to the loving overlords we have now

amateurs killing a few million people

wtf is that

[fervent thunder]

Idk if thats the most effective way to go about ut

[trim magnet]

i think u been smoking too many georgia stones

[fervent thunder]

Maybe make 7.5 billion inject poison while brainwashing them that it is a cure for some disease they made up.

[low moon]

nah man thast what it said

word for word

in multiple languages ot make sure we dont missunderstand

:))

[fervent thunder]

🤣true

[trim magnet]

yea a manmade stone that was built in the 80s

[low moon]

well ask yourself would you build that. who woudl build it

its just a prank huh

and now it got bombed

weird stuff

[trim magnet]

i can build something like that and bury it to troll some qanons like yourself easy

[fervent thunder]

Then watch wef

Where most politicians gather

Of eu and us

And word for word

[low moon]

I dont think people build such monuments to troll

[trim magnet]

politicians corrupt whats new

[fervent thunder]

In that forum it was said "its better if the world population was 500 million"

[low moon]

if you knew stuff and couldnt explcicitly say it whats the next best thing: architecture

[fallen drum]

4080 super vs 3090 ?

[charred mesa]

teacher reading my essay on how to combat overpopulation:

[fervent thunder]

Symbols next

[trim magnet]

a 4chan forum and name myself Q

[fervent thunder]

Have u seen "watch out for 666"?

[trim magnet]

heard they are reptilian aliens and they are in possesion of gods ark and no one knows this information ( only me and alex jones)

[fervent thunder]

Thats one way to lower the credibility of "conspiracies", by spreading crazy ones too.

So people associate plausible / true conspiracies with crazy.

Alex Jones (controlled opposition) does a great job at that

[trim magnet]

that may be true,they use those ppl to paint ppl who oppose wars as loons

[nova zodiac]

You got a billion dollar debt hanging round your neck too?? 😛

[wet dawn]

hi

[grizzled palm]

Hi! I'm an artist too. 😊 (Pencil and paper kind, with books on figure drawing and color theory on my shelf.) I'm working on a painting app that has AI features built in, with a feel similar to Procreate. Would you be interested in experimenting with it? Also looking for help making tutorials / use videos.
https://github.com/QuintessentialForms/ParrotLUX

[supple pond]

hello

[bleak matrix]

Please check our rules regarding advertisements! (Also, good morning, and I hope everyone is well today!)

[bleak matrix]

You might want to check out #1092446741984444416

Love the concept!

[fervent thunder]

/D

[empty patrol]

gm

[ornate lily]

hello

[low moon]

Spinning is a good trick.

[versed goblet]

hello

你们好啊

[wicked bolt]

hello

[grizzled palm]

Did not know about this section. Doesn't look like people visit it, unfortunately. 😕

[trim path]

good

[hybrid gazelle]

Hi, everyone, can someone help me with generating a qr code with an image inside?😅 Would highly appreciate it

[fervent thunder]

Youtube .. qr control net

[hybrid gazelle]

I know, maybe someone did it already, I need to try how the thing will look merged with the logo

[dapper wasp]

hello,youtube

[bleak matrix]

We've had it for quite some time! I always like coming back to it and checking and seeing what everyone else is up to. What are the plans for your project?

[rotund igloo]

1242I know

[grizzled palm]

Posted a summary in the community section. I'm mostly making the app to use it. I do want it to be useful for other artists though, so if anyone's willing to give me feedback, I can improve usecases I haven't thought of. If you know any digital / traditional artists interested in AI who might be interested, please feel free to point them my way.
(One kind user here needed to do some stuff with IP redirecting, and by fixing the app for his use case, it's now even more general and powerful. 🙂 )

[rotund igloo]

hello everyone!

[vital sky]

I can't wait for stable diffusion 3 to come out

[keen swan]

By any chance is the v2 beta API using SD3? It seems the quality is better for some reason than stable-diffusion-xl-1024-v1-0

[opaque gate]

We've had it for quite some time! I always like coming back to it and checking and seeing what everyone else is up to. What are the plans for your project?

hi

[fair pewter]

animediff

on comfyui

[orchid vault]

heklo

[fair pewter]

hey ken kanaki

[bleak matrix]

That's awesome!

If you want to make tools for artists, I suggest that you look at various art programs, and study those features.

Most artists want:

• Layers (and the ability to group them)
• Clipping, masking, and alpha layers are all essential
• Tools like the ability to select, copy, paste, rotate, and transform
• With transform, you want the ability to warp, change the perspective, (which you can do with the above)
• Most artists are very into brushes--so the more tablets you support, the easier it will be for you. Having a dynamic brush system is very good. You can look at programs like CSP, Paintstorm Studio, Krita, and PS, etc. Sizing, opacity, and the ability to place in your own image, etc, are important. (Esp cuz going from AI to brush would be useful.)
• You also want bucket fill tools, and be able to drop textures in/out

Those would be my suggestions for you, as a base.

[grizzled palm]

Thanks! I've already finished 13 out of the 17 features you listed. I don't technically have custom brushes, but they do exist (defined in JSON files w/ brush tip + texture images). Will add a way to customize them in-app soon.
Missing 3 things:
No clipping layers. (What are those. Have to look up. Sounds redundant with masks+alpha?)
No select-cut (and no plan to add... but it wouldn't be hard tbh.)
No perspective warp. Hmm. That would be super easy to add. The layers are already 3d quads. (Would people really want it though? Why? 🤔)

To brag a little: my flood-fill tool is top-tier! You can expand the flood with padding, so there's no gap against the lineart! 🙂 (But it's also too slow, like 2.5 seconds. Moving it to the GPU is on my todolist.)

I also have not just best-in-class, but only-in-class features: Like the ability to freely arrange layers of different resolutions. E.g., you can have a 1024x1024 layer with your entire image, and then another 1024x1024 layer that overlaps that layer, but smaller, only covering a face, so you can inpaint and have a hi-res face. (You can merge down and lose the hi-res, or just keep it. Export at whatever resolution makes sense for the whole composition.)
(Also best-/only-in-class: general/universal AI API compatibility that feels like native in-app integration. Also: intuitive + non-destructive inpainting & upscaling, so you can keep inpainting different parts before & after upscaling etc. Also: intuitive controlnet and img2img configuring with node-links. And other stuff. 😉 )

[jade rivet]

hi everyone

[clever hemlock]

hi

[slow aspen]

hello wans

[hoary ridge]

every one let here

[neat hearth]

hello

[forest wyvern]

hello

what day of today

[trim magnet]

the day when you invite your 6 bots to the server

[abstract stratus]

ewgdvheard they are reptilian aliens and they are in possesion of gods ark and no one knows this information ( only me and alex jones)

[trim magnet]

ewgdvheard

[spice crypt]

heard they are reptilian aliens

[tranquil rose]

good day, chaps!

[tepid vortex]

hi

[zinc glen]

你们好啊

[cyan quarry]

hello

[fleet pelican]

你们好啊晚上

[tranquil rose]

jestryyyyy

wazzup

[dense parcel]

你们哈皮！

[topaz parcel]

Question: Does anyone know how i can set the generation of an image2image using AUTO1111 to stop at say 90%. It seems to get to about 90% and I like it, but then for the last 10% then 'overcooks it'. I have tried lowering the denoise and CFG. But still that last 10% it adds way too much detail, but is perfect at 90%. How can i make it automatically stop the img2img generation when it gets to that point?

[subtle egret]

晚上好！

[candid nimbus]

good evening

[amber bear]

GOOD EVENING

[tidal tangle]

hello\

[amber bear]

does anyone know if they're going to put SD3 in here for testing like in SDXL?
I'm not necessarily for it, I thought the resizing of the images made SDXL preferenced towards really large subjects, as an affect of the voting

to an extent

@bleak matrix can you tell Emad (or whoever runs show now?) lol

[paper leaf]

hello

[static cape]

Funny... just wanted to write that SD3's API access should finally start this week... it's mid April. ^_^

[sudden ruin]

There already is a testing bot for SD3 with voting again, just not publicly available

[thick cliff]

Thank you! I'm using them to test PAG. https://civitai.com/posts/2163684 (credits in the description)

[fallow fog]

Seems amazing ^^

[amber bear]

Don't you think they should require full sized images during voting?
I mean there is a huge difference between an image iconized vs. as something that's y'know, 256^2 and up. So the (312ish?) sized images are going to be best at that size if the voting is at that size--when normally they're at something like 512-1024 or more, in probably 80% of display situations.

Or do they compensate for that minimization/resizing affect somehow or another?

[compact swan]

Hey, does anyone know a cheap way to use SD on a phone with lots of free credits?

[jade jackal]

setup a local sd on a pc with a public Url and authenticaiton and use that to login on the phone?

[dry trellis]

Not sure if it will be blurry but you can right click the preview window when it's at 90 percent and save it

[karmic cedar]

I’m ready to generate fish using Stable Diffusion 3 Alpha Turbo

I will make the most fish

None of you are ready for it

[dry trellis]

click "Generate Forever" and actually let it run forever

[fervent thunder]

Im looking for 3d to anime while keeping the same character, what should i used, text to image, image to image, and what control nets?

also what can be used to keep the colors more or less the same ?

[dry trellis]

The Reference Only controlnet should keep the colors similar

[fervent thunder]

thx will try it

[clear oyster]

why does it take like 10 minutes (seriously 10 whole minutes) for sd on cpu?

is this normal?

anyway to speed it up with commandline arg arguments?

[dry trellis]

I think that's normal for cpu

[fervent thunder]

yeah, took like 4 min even with 1660ti, thats why i spent 1400$ on pc build, now its 1s per img at 1080p

[charred mesa]

pixart sigma is a great substitute for the lack of SD3

[clear oyster]

what gpu gets 1s per image?

[fervent thunder]

3090

[clear oyster]

nice!!

[fervent thunder]

thx loving it so far

[clear oyster]

turboxl or standard sd?

[fervent thunder]

sd xl

[clear oyster]

i see

[fervent thunder]

experimenting a lot

ponyxl atm

[sage reef]

wait so for pixart sigma i need 20gb for just the T5 part, are you kidding me?

[clear oyster]

if I do 4 image batch then the combined time is around 4 seconds on GPU

I was using CPU so that I can run llama 2 in gpu with sd but appanrety that wont work lol

[fervent thunder]

what gpu u got

Recently i tried command r on gpu and it used all 24gb vram lol, but it runs fine. 1 year from now surely 48gb vram will be the new standard

[clear oyster]

just a lil 4050 (6 gb vram)

[clear oyster]

command r is what??

[hybrid gazelle]

Can someone generate a qr with logo for me?😂

[fervent thunder]

llm, like llama 2, but better, gpt-4 level and open source

[clear oyster]

oh I see

[fervent thunder]

also uncensored

can give you spicy stuff, both nsfw and not so legal ones

[clear oyster]

looool interesting

must not be based off llama?

[fervent thunder]

ofc not

[clear oyster]

last i checked every single open source llm was based off llama

[fervent thunder]

yeah llama is woke af

this thing is uncensored, it will literally teach you to cook coke

[clear oyster]

eh I mean not really, its same as all the other llms imho (still woke for sure)

[clear oyster]

LMFAOO well thats quite uncensored

i imagine the system reqs would be higher

[fervent thunder]

its 104B parameter model, i was surprised 24gb vram was enough

[jade jackal]

if you are using HiRes Fix... that blurryness at 90% happens if you setup odd sizes, change the upscaler from Latent to None (or other upscalers)

[clear oyster]

104b?! damn yeah wonder how 24gb worked lol

[fervent thunder]

gpt 4 is 1.7 trillion parameters, its a wonder how in just 1 year 104B parameter model has same performance

[clear oyster]

yeah thats also true

[dry trellis]

I wonder if it can tell you how to make coke the soda? like figure out the secret ingredients 😄

[clear oyster]

llama3 should surpass gpt3 in just 7b
launch maybe june

[clear oyster]

oh tha would be nice hahat

anyone here has hands on enterprise GPUs?

[warm bane]

Guys, what is the AI that generates stuff to fill an image

I have a vertical image character and I want to fill the sides with a suitable background, how do I do that?

[clear oyster]

hmm, same question

[dry trellis]

You can use outpainting for that

I use Fooocus outpainting

[clear oyster]

i should use "--lowvram" to lower vram usage right?

with lowvram flag only .4 gb vram is used lol

gets around 4 s/it though rip lol

[zinc epoch]

hello guys

[clear oyster]

how do you unload a model?

[zinc epoch]

听不懂

兄弟

能不能说中文

[clear oyster]

"Can't understand
brother
Can you speak Chinese?" - ^^

no lol

[charred mesa]

idk in 8-bit doesn't go over 12GB iirc, but I need to test it again

people blow this T5 thing out of proportion

[fervent thunder]

can i have a custom shaped object and outpaint on that, for instance a flower with no background, and generate an image on someone holding it?

[shell tendon]

it doesn't need even close to that for vram

[charred mesa]

I'll test it in T5 in 4-bit

[dry trellis]

Yeah just have to be careful painting around it

[clear oyster]

what are the flags that makes sd run faster?

[charred mesa]

@sage reef Pixart with T5 loaded in 4-bit, it doesn't go above 8.6GB VRAM

[fervent thunder]

have u tried forge

[charred mesa]

you can also just load it on CPU and stuff

at 8-bit it's around 10.3GB

[clear oyster]

whats that?

[fervent thunder]

try it , it makes sd work 2x faster on low vram pc
https://github.com/lllyasviel/stable-diffusion-webui-forge

[zinc epoch]

whats what?

how about start a League of Legends, im waiting.

hello

[clear oyster]

If you use less powerful GPU like 6GB vram, you can expect to get about 60~75% speed up in inference speed (it/s), the GPU memory peak (in task manager) will drop about 800MB to 1.5GB, the maximum diffusion resolution (that will not OOM) will increase about 3x, and the maximum diffusion batch size (that will not OOM) will increase about 4x.

looks nice

[zinc epoch]

how about i provide gpu, because im in china, i know where can find the cheap card

[sage reef]

but wait, are you using comfy or what? i would love to try it inside comfy

[charred mesa]

yes

[zinc epoch]

Any guys interested?

[charred mesa]

if you have 12GB of vram and you load it in bnb8bit then I think it should be fine

[sage reef]

kk

[fervent thunder]

i feel like u need a full time job using sd to learn comfy

[charred mesa]

https://github.com/city96/ComfyUI_ExtraModels

its avaialble in the comfyui manager

[sage reef]

nah comfy is ez :3

[charred mesa]

I recommend using extrasamplers -> res_momentumized (or whatever) + cfg_rescaling with like 0.8

[clear oyster]

jsut the defauly comfy workflow with the bare minumum is easier than a111

[sage reef]

i tried sigma on the demo huggingface, some results look kinda baked

[charred mesa]

yeah cfg rescaling helps with that too

and whatever the sampler is

20-30 steps is okay with res_momentumized

[sage reef]

yea il try some combinations, thx for info

[charred mesa]

you're welcome

its one of the slower samplers though

like DPMPP SDE and Heun

[sage reef]

yea i think i tried that sampler with cascade, was slow if i recall

oh so im looking at the extramodels thingy, so for T5 we can offload the whole thing maybe on the cpu side, since it's literally just a text conditioner, dont have to trouble my gpu with that then, and then use gpu on sigma+vae+whatever else

[charred mesa]

on cpu it takes a bunch of RAM

[sage reef]

hmm il experiment with a combo that works

i wish there was a non diffusers version

maybe even pruned version

but wait, i already have a t5 from ella i think, can i use that one? or is this drastically different?

[charred mesa]

ehhhh idk

[sage reef]

oh someone asked the question im asking already lol https://github.com/city96/ComfyUI_ExtraModels/issues/20

[shell tendon]

res is def a relatively slow sampler, but man, it cannot be replaced

#🏞｜general-with-images message really great with this denoise schedule

[sage reef]

are there any custom nodes that allow us to run Kandinsky models inside comfy?

i never tried those

[jade jackal]

so i learned how to use those dynamic prompts... and magically its like my hard drive space is gone...

[topaz parcel]

what are dynamic prompts

[trail lion]

a {tall|short} {man|woman}, produces "a tall man", "a short woman", "a tall woman" etc, you can do quite a bit with it

think of like hair colors, lengths, styles...hard to imagine what life was like before dynamic prompts

[pearl ocean]

Spotify going up another dollar

A$13.99/month

[trail lion]

yah I have the family plan, my daughter insisted on having spotify, I used to have pandora which was way cheaper

if you have ever generated regularization images for training, excellent use case for dynamic prompts, to get a nice variation in your data set

[topaz parcel]

when did these come out - will give it a try

[analog orbit]

https://www.reddit.com/r/StableDiffusion/comments/1c50m7z/roopfacefusion_may_be_unsafe_binaries_have_been/

[shell tendon]

wow

huge red flag there

[valid echo]

are sdxl loras compatible with pony?

[trail lion]

yup, just grab the extension and you're good to go

[coral quest]

早上好！

[glass frost]

hello everyone

how to begin?

[last wasp]

i am come

[patent prawn]

how wo begin

everybody

[loud gale]

hi bosy

[dry shard]

yeah

[karmic cedar]

hi are we ready to start?

[trim magnet]

yes we need more february bot accs then we ready

[karmic cedar]

Alright it’s 7:31, we are going to begin our stable diffusion 3 lesson everyone

everyone get ready to open your Internet Web Sight Browsers

now i want you to remember your first childhood memory

[sage reef]

say whut cuh

[karmic cedar]

now that you’ve recalled your first childhood memory, go ahead and type www.stablediffusionthree.com

while you wait for it to load, please try to remember the first time you got a boner

or, if you fancy, a chub

moving on, if the website doesn’t resolve, i’m afraid you’ll just have to wait until Stable the Clown announces the official release of SD3

[sage reef]

[trim magnet]

gm,how to use,hello bossy,how to begin 🤖

how to use

[sage reef]

welcome my friends, let me tell you how to use

[edgy tundra]

how to use it

[trim magnet]

how to use it

[turbid shard]

Can someone help me with this error? AttributeError: 'NoneType' object has no attribute 'lowvram'

[sage reef]

you are using it wrong

[turbid shard]

how so

[sage reef]

but wait, where did you get that error, what were you doing?

[turbid shard]

trying to load a model

[sage reef]

comfy?

[turbid shard]

SD

[sage reef]

yea but you using comfyui or a1111 or what?

[turbid shard]

a111

[karmic cedar]

how can be

???

[novel blade]

how to use to download

[trim magnet]

how to use to download

[sage reef]

oh i dont have a1111, so dont think i will be able to help... hmm, can you load other models fine or is this your first time with a1111?
does your pc have enough specs to run models? @turbid shard i mean i did use a1111 before but never had problem loading models

[fervent thunder]

If I'm interested in doing mostly anime generations, do you guys reccomend SDXL or SD1.5?

[pearl ocean]

XL

[sage reef]

xl has some cool anime models

[fervent thunder]

what models do you reccomend for it?

[turbid shard]

can't load anymodel

[trim magnet]

both works,u just get less deformities with XL

[sage reef]

but what are your pc specs?

just tell me vram, ram

[turbid shard]

I have enough vram, specs are not the problem in my opinion

[sage reef]

well can you quickly check if you can load in comfyui or another interface instead of a1111, we need to rule out specs for sure :3

[trim magnet]

but whats the gpu

[turbid shard]

im not even try to generate, I just can't load or change models

[trim magnet]

yes but webui uses your gpu to move/load the models

[shell tendon]

how much vram do you have?

there is no such thing as enough vram lol

[obsidian bane]

good morning

[turbid shard]

12gb

[sage reef]

@fervent thunder Animagine XL 3 , Anime Art Diffusion XL, are mainly for sdxl i use... for 1.5, i like the following: anime screencap 1.0, anything else 4.5, any lora clean anime mix (aam any anime mix, etc), kantanmix, kimix am...

[shell tendon]

i had problems sometimes with a1111 with a 12gb 3080... usually it was fine but it freaked out after a while at any batch size greater than 3, then anything greater than 1, then sometimes it'd just OOM for the hell of it

use forge, stableswarm, or comfyui

[sage reef]

yea can you just try to load inside comfy or anything else quickly, we need to see what kind of problem this is

[turbid shard]

can you use comfui with an amd gpu?

[fervent thunder]

thanks for the recs, and is 6gb vram enough for sdxl

[sage reef]

6gb seems too on the edge... ugh.. not sure... i know 8gb works for sure (even cascade)

i mean you can try :3

[fervent thunder]

what would happen if it doesnt work

[sage reef]

and remember, you dont need refiner

what would happen? i mean... it doesnt work :3

[fervent thunder]

alrighty ill try lol

[shell tendon]

yeah with comfy i'm usually at 8.5gb or so with sdxl

[sage reef]

im personally on nvidia, so no idea 😦

comfy is just too good baby ❤️

man where is sd3 already :3

my guess is around april 26

[pearl ocean]

Somewhere in the void of space

[sage reef]

nuuuu

[shell tendon]

might want to specify the year 😛

i'm going with '25

[sage reef]

haha

[forest trout]

In your opinion, which of their projects is coming out next?

[karmic cedar]

probably SD ‘97

[dapper ocean]

new

[trim magnet]

new

[karmic cedar]

old

41 on the 19th

[sage reef]

[pearl ocean]

im going with 2025

[nova zodiac]

you can - works better on linux with RoCM than it does with zluda on Windows, but it does work

there are pinned posts in #🤝｜tech-support for install instructions

[pearl ocean]

https://tenor.com/view/developer-recruiters-programmer-frontend-gif-23808695

[winged jewel]

Hey Guys, I am a bit in dilema since i have to submit the requirments to a client, and need community's help in this!
I am running a workflow wehere i am taking a input from a webcam and then transofmring it into a picture with a spcific output using cutom trained lora. These output will showed in real time when the person stands infront of the webcam onto a screen.
Now my pC which has 4090 give me each image with a delay of about 2 sec.
And now the client says he is okay with 1 sec delay, so i am thinking would the RTX A6000 ADA will be okay for this since it has 48GB vram ?
Edit - i am running a turbo and sometimes lighitning SDXL model with Depth Control net

[karmic cedar]

h e l l o

what’s diffusing

[shell tendon]

a6000 is slower than a 4090

if you want something faster you might want to just pick up a H100

[winged jewel]

What about the ADA version ?

[shell tendon]

slower

https://www.amazon.com/NVIDIA-Graphics-Memory-Learning-Compute/dp/B0CMJCD2HG?source=ps-sl-shoppingads-lpcontext&ref_=fplfs&psc=1&smid=AYCADGG1WXC5Q#customerReviews just get one of these lol

[winged jewel]

Also how difficult is it to setup a H100 or A100 gpu ?

[shell tendon]

afaik it's the only thing that's flat out faster

not hard so long as you got $30k to burn

[winged jewel]

money is not issue

[shell tendon]

well, buy a H100

crazy your client is ready to flush that kinda dough on a webcam feed lol

[winged jewel]

not a web cam feed, its a real time camera feed, i am just testing it now on my webcam

[trim magnet]

RTX 6000 ADA is faster than 4090

[winged jewel]

although the camera would be high res i would still resize the input to the normal sdxl version

I am confused now 2 people and both saying different things

[trim magnet]

it is faster u can check it yourself on vlad benchmark site

[shell tendon]

yeah this guy is right i goofed that up

ada is faster, has slightly more cores... nothing massive though

so if you need to cut your time in half, that won't solve the problem

[trim magnet]

now if u want the best of the best u should buy an A100 but SXM4 not pcie,but SXM servers are expensive

[winged jewel]

I am just worried about the post purchase installation and server setup is going to be nasty and pain in the ass lol

[bitter iris]

heello

[winged jewel]

cause i built the 4090 build and i know what i had go through ;,)

[sage reef]

why is everyone saying hello

[trim magnet]

bots check the acc creation date,all february

[sage reef]

lol

[trim magnet]

well if u gonna buy a SXM server with several A100's price is like 150k usd or more so u have to worry about that as well 🗿

[sage reef]

bye

[winged jewel]

i am not doing the purchase, client is ;), my job is to make sure the workflow is running properly, which right now is, but with the new setup it should be more faster!

basically at half the iteration at what it is right now

[sage reef]

can your client buy me a 4090? :3 it's probably peanuts to him

[winged jewel]

i hope to get 1 out of this project!

[sage reef]

i need to upgrade pc soon

but wonder if i should wait for 5090

[trim magnet]

well if money is not a problem he could probably buy something similar to this https://www.amazon.com/Supermicro-Customized-Platinum-Baseboard-Analytics/dp/B0CSVN4YXW

[thorny echo]

I just watched a 60 second tutorial on how to install stable diffusion but when running it the windows blue thing pops up saying it might be a risk which makes scene as this is open source but the file size is also only 80mb and I watched a piece of another tutorial and his one was 6gb and from a different person on github . Can someone help me find the official link? Thanks! 👍

[winged jewel]

i dont think comfyui or any sdxl model supports more than 1 gpu at 1 instance

so i think only 1 gpu is more than enough

[sage reef]

what did you install?

[trim magnet]

oh well you buy these kind of stuff for training models,so u load 1 model per gpu

[thorny echo]

nothing yet just unzipped it and opened the update exe which came up with the blue thing. It should just be the stable diffusion. https://www.youtube.com/watch?v=i5hvZvzcxoo <- thats the one I watched

[sage reef]

huh... usually the blue thing pops up when you try to open some .exe, but that is automatic1111, you use the batch script to run it, so not sure how you getting the blue thing... to what program or exe is the blue thing triggering on? cause last time i used a1111 this never happened

[thorny echo]

ok thats good that you've also used auto1111. I've also seemed to have mixed up my words as the 2 files I've tried running which are the ones giving my the blue screens are the batch files such as run and update

kinda odd as all update.bat does it call eviro.bat and pull some git stuff

[sage reef]

weird.. for a1111, i never used the prepackaged versions, i always used from source like this:
https://github.com/AUTOMATIC1111/stable-diffusion-webui#automatic-installation-on-windows

[thorny echo]

ahh okay ty I shall try it out

[sage reef]

but i mean the blue message thingy should only happen once anyway, and if it runs after, it should be fine and the message shouldnt appear again, i think. but yea that's some weird stuff...

man how many bots are there smh

[shell tendon]

wish they at least came in and spewed interesting prompts

[sage reef]

too much work :3

[fervent thunder]

Lol, thats gonna cost 3k +

[karmic cedar]

no big deal if money is no big deal

j/k money is a big deal

[shell tendon]

think waiting is a good idea right now

we got zen 5 around the corner too from AMD

[shell tendon]

doubt it

[fervent thunder]

Idk if waiting is a good idea

Theres always new gpus incoming

My the logic of wait for better prices, you would never get a gpu

[shell tendon]

agreed, but if you're building a new system, it's prolly not that great of an idea to do it right now

new cpu coming out potentially within months but highly likely this year, and a new gpu within a year

if you're the type to upgrade every 5 years not every 2, prolly makes sense to sit tight till this fall

[pearl ocean]

I just have a simple old 4080

[karmic cedar]

your uncle’s a 4080

[pearl ocean]

his dead

[karmic cedar]

😦

i rescind

[summer grail]

hi yall! I was wondering if anyone's used turboXL in deforum before? for somereason, when i use an init image of a person, and prompt the camera to move towards the left, the person's face is squished in a really weird way

[jagged mantle]

hello

[stone oxide]

One message removed from a suspended account.

[karmic cedar]

hello

[summer grail]

I think turbo is faster but lighting's quality is better

[stone oxide]

One message removed from a suspended account.

[summer grail]

Solved the issue btw! turned down the strength schedule as the frames increased in case anyone else ran into the same issue:)

[summer grail]

yes i think so

[loud solar]

Always have a look at model description on civit.ai

[summer grail]

indeed

[near vapor]

hello

[karmic stirrup]

The opening background is an office environment with a desk stacked with papers and notebooks, requiring whiteboard animation form

[small prism]

Can someone help me with a prompt template for garment generation with specific fixed attributes required in Fashion Garment industry.

[meager anvil]

hello

[deep basalt]

hello

[wanton cloak]

hi guys

[nova zodiac]

Turbo 1-2 steps, medium quality, uses Turbo scheduler; lightning 2-8 steps, high quality, uses SGM Uniform scheduler

[deep basalt]

123

news

[sturdy sierra]

123

[pine charm]

hello guys

[sturdy sierra]

hello boys

[pine charm]

sup

[nova zodiac]

superstar DJs, here we go!

[full sandal]

yes

[loud solar]

Sounds like a SL DJ 😉

[past charm]

[sweet chasm]

[rotund igloo]

hello

[frank rampart]

I wanted some help making a couples photo into a stylised ai artwork

[orchid vault]

hello

[frank rampart]

im using controlnet, RealCartoonAnime and a1111 img2img. However, the result seems to mix up various elements. Ive used the && tag to make separate descriptions of the couple however it seems to mesh the descriptions together. any tips?

[near vapor]

hello

[sudden ruin]

Lower denoise?

[sharp pumice]

cheerio

I've been trying to do a style transfer, but no luck ... I would love to be able to use img2img + LORA, but it seems this is not the way to go?

[frank rampart]

ive never messed around with denoise before. let me try it and get back to you

jea you right its getting better results with a lower denoise. its abit blurry and murky tho, let me play around and try and find the sweet spot

[sudden ruin]

Also depends on your sampler and steps, crank up the steps and lower denoise might help

[nova zodiac]

Look into ip-adapter/instant-ID

[sharp pumice]

thanks everyone ❤️

[fervent thunder]

What is the standard for character consistency?

Other than specific lora

[honest mica]

You are excluding the standard with that sentence in my opinion.

[fervent thunder]

Lora is too time consuming :/, are there other faster methods

Like IpAdapter maybe

[bleak matrix]

Good morning, everyone!

How are we all today?

[nova zodiac]

Ip-adapter, instant-ID, photomaker, roop, reActor

[nova zodiac]

We good, how u??

[autumn egret]

Did anyone who is on the waitlist for SD3 get access to it yet?

[nova zodiac]

Pretty much no

[frank rampart]

make your own lora. i followed along to this video and it got me decent results. basically use a multi expression chracter sheet in the img2img tab, with control net and adetailer. then take the results and make a lora

https://www.youtube.com/watch?v=Dn1zjeV8Tco&list=PLPAeYpPQiY11tARkMMhXFyYjjCY9u20GG&index=57

if you willing to wrestle with Koya instalation you can do it for free. or pay 10USD to civitai and you can drag and drop to make a lora

[honest mica]

Or you go the easy route and just use Onetrainer.

[fervent thunder]

Will definitely look into both. Specifically i will try train a lora based on 2d characters from movies

[tacit roost]

guys can someone help me ? where can i generate image i couldnt find the channel for it

[trim magnet]

because there isnt a channel for it

[karmic cedar]

how to use

???

[trim magnet]

miss

[sudden ruin]

You can use a mix of a few celebrity names

[fair pewter]

ip-ip2adapters

and loras ofc

[fervent thunder]

Which one gives the best results?

[charred mesa]

It's crazy that SD3 is like technically 1-3 weeks away

and yet I feel like it's still very far away

still no big wave of waitlist beta testers and etc

nothing about controlnet or current training status

[rugged mirage]

it's definitely not 1 week away lol

[bleak slate]

hey guys, i'm kinda new to SD, i've read a lot of videos / guides yet but I'm not comfortable enough finding my own solution so here's my question

I have a LOT of photos of my face and would like to generate realistic looking photos of me based on my face, I've seen boring reality on civitai but is there anyway to give it my face ? maybe make a LORA of my face ?

[rugged mirage]

if you want a quick solution start with ipadapter for faces, if you have a decent gpu straight up create a lora

[bleak slate]

i have a 3070 8Gb, should be fine enough for a lora right ?

also i'm kind of a noob but will i be able to adapt my lora in any style i want ? 🫠

[rugged mirage]

should be fine for a 1.5 Lora, not sure if you can do a sdxl lora with just 8gb tho

[bleak slate]

would you happen to have a link for some kind of step by step guide?

thank you so much for taking time to answer

[rugged mirage]

id just youtube and look for a recent guide, I havent created a Lora since sdxl came out, whatever guide Ive used then is likely suboptimal by now

[wet dawn]

hi

[rugged mirage]

just look for something like 'simple kohya lora guide' and specify 1.5 or sdxl, whichever you prefer

[bleak slate]

thank you so much

[karmic cedar]

Is kohya still with lora or did they break up?

[distant swift]

Is there any info on the SD3 release yet? I haven't checked in here in a while

[rugged mirage]

nope

not even the API yet

[sage reef]

april 26, but dont tell anyone :3

[karmic cedar]

Friday the 19th because that’s my birthday and they decided to single me out

[sage reef]

well at least it's not the 13th

[fair pewter]

both together

[distant swift]

Nah, probably April 69, which is a totally existent date

[karmic cedar]

It’s not the 16th because that’s today and I don’t have SD3

[fair pewter]

same day as my report card date

[sage reef]

welp

[fair pewter]

gotta talk to the devs and make it release on the 25th or elsei might not be able to try it out

[karmic cedar]

I think talking will help

Definitely

[sage reef]

[fair pewter]

LOL

gotta sponsor me

[distant swift]

We will probably know once it releases, so I guess there is no point in sitting here and making up release dates

[karmic cedar]

[fair pewter]

agreed

never really got the jig of sd3

what does it offer

[sage reef]

i technically didnt make up april 26, it's a decent approximation based on what was said by the new lead dude

[fair pewter]

an approx is an approx

[karmic cedar]

And then AI in the free world will officially be able to spell

[sage reef]

april 26 to may 3 is my range guess

[distant swift]

SDXL can already somewhat spell, but SD3 seems to be able to do it way better. It's a very unique model as its arch is entirely different from most diffusion models we've seen before

[karmic cedar]

yea yea it does attention guidance differently

[distant swift]

Also 2 text encoder and 1 image encoder

[karmic cedar]

^

[hot kettle]

Was in sdxl already

[karmic cedar]

we’re all very excited for the new architecture

[distant swift]

SDXL didn't have an image encoder, and didn't use T5 for one of the text encoders(which is one of the things that allow diffusion models to be more coherent)

[hot kettle]

My bad, only meant the text encoders part

[karmic cedar]

the experience of prompting will be a lot more like ELLA / 1.5

more tokens held

[sage reef]

i wish they released the image varations or reimagine xl version that they are using, cause what we have currently (even with cascade) is all based on 2.1 unclip, and that is what they used for the first reimagine, but the xl version is clearly upgraded, but we never got that (or at least i dont think we did)

[distant swift]

It did support inputs from image encoders (CLIP Vision), but due to it not being trained for image conditioning, it needed IP-Adapters to have image encoding

[hot kettle]

Afaik clip based encoder should be able to encode text and image features quite similarly