💬｜general-chat

[harsh bane]

I mean it's cool and all and I do understand how it works

well mainly because after joining I can chat here ,, all the other room won't allow me to no matter what I do , bottom line I don't know how to get in the other rooms

the room I do go to is dieing lol

but I understand ,, I'll leave if you like , no hard feelings , I am not like that

[jovial wraith]

i'm not asking anyone to leave, it's just that it's unusual for someone that doesn't have any interest in AI to join this server

[harsh bane]

AI is cool but myself that should make you aware that you are talking to AI and not a person

it's got to the point that you can't tell any more

[jovial wraith]

i knew it. i had a feeling you were an LLM hooked to a discord account. or i'm going crazy and misunderstood things

[harsh bane]

LLM?

[jovial wraith]

Large Language Model. you said in the above message that i'm talking to an AI, right? or am i misunderstanding and completely embarrassing myself here?

[harsh bane]

oh no I mint on phone calls you can tell if your talking to AI or not ,, I'm sorry

[jovial wraith]

ah, alright. apologies

[harsh bane]

but see what I mean you can't tell

[jovial wraith]

yes, exactly

[harsh bane]

phone brb

[frail void]

Hi guys

I'm coding a diffusion model on pytorch

and i need help with something

anyone down to help?

[jovial wraith]

i'm not a programmer by any means, but i'll see what i can do

[frail void]

Are you familiar with it?

welll... there's different ways to guide image generation right? with text, images, classes,etc/

i want to do it with an image

[jovial wraith]

alright. what problem are you running into?

[frail void]

How can i embed context ?

[jovial wraith]

i'm sorry, but i have no idea.

is there any reason why you want to code it yourself?

[frail void]

Yes

also is stable diffusion open source /

?

[jovial wraith]

i'm pretty sure it is!

as in, we have access to the weights to train our own models. that's all i know. i'm not an expert on this topic

[shrewd linden]

so wasn't SD3 to be released?

[jovial wraith]

we only had a vague hint which turned out to be false. we have no further info afaik

great. civit is down

[lusty beacon]

Hey

I have an issue. Where can I change the image size in stable cascade ui

it's not allowing me to make big or small images

Only 1024 x 1536

[harsh bane]

I was just given a apple tree 🙂

it's dead but that's the best kind for fire wood , besides I won't cut down a living tree

well you all have a wonderful day , nice meeting you Yellow

[jovial wraith]

you too!

[shell tendon]

"better for beginners" is a tough one... it is definitely easier to get started with A1111 but myself, i found i was really at a loss for what settings to tweak etc. until i switched to comfy. now when i come back for a specific task (using forge) i know exactly what to do.

[iron slate]

would multiple gpu's be better than one dedicated gpu, thinking of getting the AMD Radeon™ PRO W7900 48GB but dont know if a few smaller cards would work better

[shell tendon]

is the primary purpose for stable difusion? if so, nvidia should be the only thing on your list

which sucks, but it's the truth

https://www.pugetsystems.com/labs/articles/stable-diffusion-performance-professional-gpus/

[iron slate]

stable diffusion and a personal oobabooga based Bot im making with someone else

[shell tendon]

https://www.pugetsystems.com/wp-content/uploads/2022/08/Stable_Diffusion_Professional_Automatic1111.png

cool

are you planning on doing a lot of training, or just inference (image generation)?

[iron slate]

looking into having it be constantly running as a server type thing, and was thinking about running linux but that has issues with Nvidia(which is my current PC)

[iron slate]

just inference and api stuff

i dont understand how to train though that is something im looking to learn

[shell tendon]

i have a 4090 and that's def been a good experience, but it sounds like youre willing to drop more on a gpu than i was - 5k or so is fine for ya?

rtx a6000 is similar in price to the pro w7900 and much much better for SD

[iron slate]

yeah, itll take awhile to get but i dont mind spending that much, my current pc is around 4-5k

[shell tendon]

if you think you'll get seriously into finetuning models, that's prolly the card to get in that price range

i thiiiink the 4090 is a lil faster for image generation, but the a6000 has extra vram, which is handy for training (not absolutely essential, and if you're just wanting to train LORAs, 24gb vram on the 4090 is more than enough)

[iron slate]

Is it linux compatible lol, i see its nvidia(only asking cause Linux has been a PAIN to set up on my pc)

[shell tendon]

yep, nvidia runs better on linux than windows tbh

at some point soon i'm gonna build another pc that'll be headless with linux for that exact reason

[desert nebula]

I improved my generation speed 5 times for 130$ lol

[shell tendon]

how so?

[iron slate]

oh? i was told otherwise. good to know

[desert nebula]

sold 1080ti which has no tensor cores and bought 2080ti

[shell tendon]

i mean, double check with someone with an a6000 before dropping 5k on it, but that's what i've heard universally

[iron slate]

im thinking of doing that but i have spare parts so meh i can reuse without buying extra

(reply to headless linux)

[shell tendon]

there's some memory optimizations etc that are only available on linux (at least officially - triton, flash attention 2)

[shell tendon]

that's a huge step up yeah

[iron slate]

my current pc is pretty heffty ive been told but i have no concept of how good or normal it is

tbh im not much into the spec of PCs so thats why i need to ask around, still learning this stuff

[iron slate]

(AMD Ryzen 9 3900XT 12-Core Processor, 128GB of RAM, NVIDIA GeForce RTX 3080)

and Vram still maxes out when running my bot

really my whole gpu does

[shell tendon]

yeah that's not bad at all

10gb vram?

[iron slate]

yep

[shell tendon]

here's the real question for you i guess. what is more important: inference speed, or ability to fine tune large models with the convenience of doing so at home? (the latter you can offload by renting time on A100/H100s if need be)

if you see yourself primarily doing image generation and maybe occasionally tinkering with training, get the 4090

[iron slate]

i prefer all free once i set it up, which is why im going big

[iron slate]

i plan to train a bunch once i learn how

its getting there thats been hard cause i dont really understand the stuff yet

im still new-ish to AI stuff considering most people

[shell tendon]

how big of training projects is the question though

it's annoying as hell to me but you can't just spend 5k vs 2k and get something that is just better

nvidia price gouges based on vram as they know it's useful for ai and they have no real competition

their cards wipe the floor with amd right now (ugh)

so the a6000 is slower than the 4090, but... has 48gb vram instead of 24gb.

if you're looking to train a full checkpoint, you might find that 48gb handy. if you want to train LORAs - single characters, styles, objects - 24gb is more than enough.

for the difference in price, it might be wiser to get a 4090 and sock away that extra 3k for time on runpod.io

as card prices will ilkely drop over the next year or two anyway and/or get more powerful

[iron slate]

ill be using it for: Voice, Image, (Video once its out), and text bots, training wise: Pictures, Voice, and Text generation. probably stuff from grounds up

Im sorry for not being more knowledgable on this stuff, i really want to learn but its been tricky

[shell tendon]

another thing to consider: i'm not an expert by any means on LLMs (to say the least) but i believe you can split the vram on those between two cards

without the need to load eveyrthing into vram on a single card, dual 4090s would be cheaper than the a6000 and have 3x as many cuda processing cores

[iron slate]

ooooo!

im all for saving money LOL

i like going big but that sounds bigger

[shell tendon]

that would also allow you to dip your toes into it first

get one 4090, see if it's enough for ya

if not, get another, but make sure your mobo and your case etc are gonna be able to support it

that includes literally, as in physically suport it, these cards are so heavy they're prone to cracking if not supported with brackets or braces of some kind

[primal star]

wasn't sd3 launching on monday

[shell tendon]

was it?

[primal star]

that's what I read here earlier

guess there's still some hours

[shell tendon]

do you have a link?

or the quote

[iron slate]

oh yeah, ill be building the PC, so ill make sure to brace

[shell tendon]

the issue i'd be careful of is that you have the ability to brace two cards and there's sufficient space and ventilation so they don't just cook each other

and of course sufficient power from the PSU with enough cables to power it

either way yeah 10gb vram is not enough imo

[iron slate]

i mean it works, just slowly

[shell tendon]

i bought a gigabyte 4090 oc a few weeks ago, felt crazy for doing it, 3 seconds after starting the first sdxl generation it was done and i knew it was worth it lol

it displaced a 3080 12gb

[iron slate]

i need help setting sdxl up cause i see nowhere to install it

[shell tendon]

i bought that before i discovered stable diffusion, and was so glad after i discovered it that i got the 12gb model for whatever reason

[iron slate]

unless its a first come first server thing

[shell tendon]

cuz i was constantly running stuff that hit 10.8gb, 11.2gb, etc

what program are ya using

forge? comfyui?

[iron slate]

neither

im doing everything manually via windows terminal

im using the normal stable diffusion webui

[shell tendon]

automatic 1111?

[iron slate]

yeah

its the only one i can find a download for

[shell tendon]

https://github.com/lllyasviel/stable-diffusion-webui-forge

[iron slate]

probably outdated by now but im not sure

[shell tendon]

this is a lot better IME with vram and inference speed

basically the same thing but with the back end code replaced

and a lot of common extensions already included

[iron slate]

oh nice!

maybe animated diff will finally work

[primal star]

can't find it, guess I misread

pretty interesting take here though
https://fixupx.com/EMostaque/status/1763907760586916112?s=20

[barren junco]

I have a dumb question. Will SD3 work with ComfyUI out of the box?

[shell tendon]

probably within a day or two max is my guess, if not at launch

[shell tendon]

oh that is def interesting

[iron slate]

thank you btw

for the help

i might be back for more but idk

[shell tendon]

np

[iron slate]

SO MUCH FASTER

[shell tendon]

ha is it?

what version of a1111 were you running - when did you install a1111?

[iron slate]

i installed the lastest and i installed yesterday

[shell tendon]

it'll show the version at launch

ahh k

[iron slate]

im using the one click install of forge

[shell tendon]

yeah someone here was telling me early this AM that the newer versions of automatic 1111 fixed the speed issues... i was pretty skeptical

glad forge has improved things for ya

[iron slate]

it has by a good amount

probably trippled my speeds

[shell tendon]

wonderful!

and yeah to use SDXL just donwmload a SDXL model and drop it in your models/stable-diffusion folder

https://civitai.com/models/133005?modelVersionId=348913 this is usually the first one ppl go to use

[iron slate]

so pretty!

[jovial wraith]

i second this! JuggernautXL is a phenomenal model!

[sudden ruin]

Tell your friend that this isnt exactly the right server to ask such stuff

[iron slate]

ahh true lol, and i actually do mean a friend, i do the share thing. but yeah wont ask that again my bad

slipped my mind sorry

[sudden ruin]

No worries, no one harmed or anything

[iron slate]

oh yeah is there any way to make Stable diffusion auto input a negative prompt?

[final garnet]

Does anyone know where to get the Euler a Karras sampler? I've been searching and cant seem to find where to download the sampler or which UI comes with it.

[jovial wraith]

what ui are you using? afaik all uis ship with them preinstalled

[final garnet]

thats what i thought too. I've tried a1111, comfy, and forge trying to find it. I can see it in png info from some images, but cant find where to actually use the sampler.

ahhhhhhh, okay i understand now. looked closer at the images that had euler a karras. they were in Comfy UI, but in order to get "euler a karras" you set the sampler to euler a and the scheduler to Karras.

[jovial wraith]

ah, glad you figured it out!

[mortal delta]

are there any ways of getting the same charater without a lora or a character but with a different expression or pose? maybe even diffrent outfit?
i would like to try to make something using ai assets and such.

[zenith frigate]

Hi, how can I find out how much SDXL 0.9 would cost on the Developer Platform for 1024x1024 50 steps?

I can see the XL 1.0, but not the price for the other Developer Platform models (here: https://platform.stability.ai/pricing).

[broken cave]

what is the idea?

why would you use sdxl 0.9?

[zenith frigate]

I build and AI app that can generate images, and I include every interesting models, do you think the 0.9 is not worth listing if i already have the XL 1.0 ?

[zenith frigate]

I was wondering the same thing about SD 2.1,
it's part of the list of available engines on the Dev Platform, but yet it's not listed on the pricing page, is it not so good?

Also, does anyone know if Stable Cascade will be added to the Dev Platform?

[hexed chasm]

Sorry for ignorant question - but is Stacle Cascade an official checkpoint release?

[zenith frigate]

I think so, I've just saw this blog post https://stability.ai/news/introducing-stable-cascade

[oblique edge]

I've noticed when launching my stable console it says cuda stream: false.... this normal?

[fervent thunder]

Does anyone have a sense of when the SD 3.0 model will be available in "beta" (by which I mean, running it remotely, without a local release available initially)?

[oblique edge]

all you can do is join the wait list

I'm assuming the waitlist is just for people running locally

if you get 2 running i THINK, someone correct me if I'm wrong, but you'll just put the .ckpt file in your stable environment once it's released

[fervent thunder]

My understanding is that the waitlist is to run it remotely (since, once one can run it locally, the сat is out of the bag, so-to-speak). If they offer it remote-only, they'll have more control, such that nefarious persons can't do sketchy things or immediately create modifications of the model.

[oblique edge]

you might be right, I figured the waitlist was FOR those able to run it locally

[fervent thunder]

I think, before they release the CKPT, they are going to do something like MidJourney, where they insert themselves between you and the model.

[oblique edge]

nah

doubtful

that sorta defeats the purpose of open source

but like I said I can't say anything with certainty

[fervent thunder]

I agree. I hope you're right, and they'll just release the CKPT, but then...what would be the point of a waitlist? Limiting download bandwith on their website?

[oblique edge]

They're letting people super skilled with generative AI test it first

I mean it will have bugs

so they're going to want to release something as clean as they can

I'm hoping it's out this month, but I'm keeping my timeline releastic so april/may?

[lost quail]

sup

[fervent thunder]

Fair enough. I've developed some custom embeddings and such, and worked quite extensively with other tools (DFL, Wav2Lip, VideoReTalker, etc.), but I'm not some master GitHub developer with a million stars on my profile or somewhat.

[fervent thunder]

Sketch

[lost quail]

what is bro yappin about?

[oblique edge]

If you know you know lol

[lost quail]

for sure

[fervent thunder]

Primarily, about what sort of person(s) SD wants to have beta test the upcoming SD 3

[lost quail]

okay

[fervent thunder]

Also, StableDiffusion is on that R. Kelly math, that Microsoft math...

3.0, 3.1, 95, 98, 2000, 7, 8, 10, 11...

We have SD 1.0, then 1.2, then XL, then 3...

[oblique edge]

If you have any creative prompts that work on other models DON"T share them here, I think we have a rat. I'

[lost quail]

So some nerdy stuff

[oblique edge]

Probably coincidence but I shared a prompt here that I was using to have Gemini bypass copyright protection and people generation. Got zapped 30 min later

[lost quail]

what is this server about?

[oblique edge]

I made another prompt almost as wacky that SHOULD have worked. Well it did work, after my prompt, I asked it for bugs bunny in a dress. Got this.....

[fervent thunder]

Primarily, this server is populated by persons whom'st've's'd'nt yet been banned for shilling crypto.

[lost quail]

oh okay

[fervent thunder]

Man, I need to set a profile image.

[lost quail]

me too

but I want to stay lowkey

[oblique edge]

I made another prompt almost as wacky that SHOULD have worked. Well it did work, after my prompt, I asked it for bugs bunny in a dress. Got this.....

I put it in show n tell

[fervent thunder]

Wasn't AI used to frame Roger Rabbit?

[oblique edge]

but you can tell they did something

[oblique edge]

I see what you did there

[lost quail]

bro

I joined here

because I was looking for some big server

and I dont have to verify my phone

because I aint doin this

[fervent thunder]

Like I said, sketch. Now, I'm also sketch. It takes one to know one.

[lost quail]

wdym by "sketch"?

[fervent thunder]

"Sketch", as in, if you wrote my username on a blackboard and then sketched over the underscores.

[oblique edge]

you joined here because you were looking for "some big server". My question is some big server for what?

[lost quail]

alr

[lost quail]

for big community

to not be bored

waiting a lot of time for respond

[oblique edge]

Well this is the stable diffusion dev discord

[lost quail]

oh

wait

[fervent thunder]

I don't use Discord much but I am quite framiliar with Telegram and the dominant form of SPAM there is for crypto projects, though I assume there's plenty of variety of SPAM.

[lost quail]

I heard about stable diffusion

somewhere else

[oblique edge]

you should check out open-ai art discord

[lost quail]

oh yeah I remember

[oblique edge]

limewire

[lost quail]

rule 34

I know stable diffusion from rule 34

also hi res

midjourney

etc

[oblique edge]

I dunno what he's looking for lol

[lost quail]

rule 34 is a porn web side

[fervent thunder]

Is talking about CivitАi generally permitted? I mainly use modified models from there. I don't do porn though.

[lost quail]

basicly some people drawing some delusional-porn stuff

or more like delusional people drawing soemthin

[fervent thunder]

Bigger server = more people who will read advertisements for SuperShitGoingToTheMoonToken

[lost quail]

I was also wondering about Opera GX

but they tell me to verify my phone

[fervent thunder]

Do you know the meme image of "Mexican boy holding a Crucifix"? Create a custom keybinding for it, because you're gonna need it.

[lost quail]

I thought that was a black guy

bro I aint clickin this

after I clicked some link on discord

[fervent thunder]

The details are irrelevant; the key is, you will need to be able to invoke supernatural protection against sketchy people.

[lost quail]

it made send a lot of nsfw server invites to many ppl

from my acc

[lost quail]

bro are you high or what?

[fervent thunder]

No.

[lost quail]

or you really into nerdy stuff?

[fervent thunder]

So, I was in an AI model-training-related chat. Some dude started talking about how he was having trouble finding good training data because "women don't naturally grow cow ears". I called upon the Alpha and Omega, the God of Abraham and Issac, in the form of a Mexican child holding a crucifix, to protect me from the AI coomers.

[lost quail]

dam

theres a lot of AI coomers I guess

[brave vigil]

I'm having a terrible time training on this style.

[lost quail]

but idk why you want to call God bc of this

[brave vigil]

I built dataset of 50 decent images. Good resolution, some headshots, some body shots, some action shots with multiple people - even a handful of "detailed background" images

And the style is consistent

I captioned them as best I could, and I've tried a whole bunch of different configurations, but the results are inconsistent and rather poor

Any advice?

[fervent thunder]

What style, specifically?

[oblique edge]

Triple your data set and see if it makes a difference

if you double/triple it you should see noticeable results, if not I'd say you're doing something wrong

[fervent thunder]

I call upon the LORD to defend me against AI coomers because "artificial intelligence", in extremis, is a graven-image factory.

[brave vigil]

#🏞｜general-with-images message

As an example

[lost quail]

alr

[fervent thunder]

Are you trying to train a subject (specific person) or a style?

[lost quail]

I also wanted to create some nsfw stuff on some ai webside

but I had to pay something

it was only 10 bucks but I am not really into paying by internet

[oblique edge]

It's okay, trains leaving the station whether the boomers are on it or not

[lost quail]

like nsfw stuff was for money

[brave vigil]

On a style. I've isolated a set of prompts which give me really nice traditional illustration style art. Well, about half the images are good. I want to wrap it all up into a LoRA so I can produce it more consistent

So I made 50 quality images with variety, captioned them, and now I'm training as best I can

[oblique edge]

I haven't started messing with training but what I said should work

[brave vigil]

Guhhhhhhh

[oblique edge]

at least to give you an idea

[brave vigil]

It took a whole day to build the first 50 images. Tripling it is going to suck

[oblique edge]

dude

[fervent thunder]

@brave vigil So, the first thing I would say is, "the easiest way to complete a task is to have someone else do it for you". I would check on CivitAI for a foundation model that is as similar to the output you wannt as possible, and thenn train a LoRA from there.

[brave vigil]

Haaaaaaa!

[fervent thunder]

Are you going with a LoRA or Hypernetwork? Custom Embeddings work well enough for subjects (ex. training a specific person) but not well for styles.\

[brave vigil]

https://tenor.com/view/its-so-simple-and-yet-genius-herbert-garrison-south-park-s5e11-the-entity-gif-21984116

[brave vigil]

LoRA

[oblique edge]

I'm gonna back out of this convo now talk to Dani lol he's way more knowledgeable

[brave vigil]

Still havn't really figured out what a hypernetwork is, and how if differs from an embedding

[fervent thunder]

Hmm, 512x512 or 768x768 native?

[brave vigil]

I appreciate the help!

Mostly 768x768, some 768x1024 and some 1024x768

Headshots, body shots, shots with two or more people, shots with detailed background

[fervent thunder]

Honestly, even though SDXL V2 is "better" and has 768x768 native, my rule of thumb is, "resolution (linear) has to at least double to see a difference", so I tend to still like to work with the 512x native forms.

[brave vigil]

Different body types, age groups

[oblique edge]

So I joined here a few days ago and have been using Gemini professional (free for 2 months right now) to teach me the different terms I see people throwing around on here, and teach me how to use them in my local SD2

[naive thorn]

are there people who already own stable diffusion 3? or stability ai not giving it away yet

[oblique edge]

It's a really effective system

[fervent thunder]

@brave vigil So, the first note I'd say is, all the training data should be in the same resolution, which should match the resolution of the base model (which is almost certainly either 512x512 or 768x768). Otherwise, auto-cropping will occur, and it's usually bad. There is AI-assisted cropping available, but honestly with <100 images, you should just do it yourself in PhotoShop. Do your images contain the EurION Constellation (very important)?

[oblique edge]

There's a waitlist

[naive thorn]

i know ...............

[oblique edge]

So it's out for some

[naive thorn]

i hope it gets out at least this month

cant wait to it

[oblique edge]

I joined I'm hoping to here something from them in the next few weeks

[brave vigil]

I don't know what that is. I'm using the FurryBlend checkpoint (link below) which is itself built of SD1.5

[oblique edge]

I've got plenty to keep me busy immersing myself in 2 so it can come whenever

[naive thorn]

also i hope my pc can run it

[brave vigil]

REDACTED

[oblique edge]

lol

[naive thorn]

why they made it with waitlist this time?

[brave vigil]

I've tried other models - more generally accepted models - but this one has given me much better results

[naive thorn]

i do not remember they published a model with waitlist

[brave vigil]

Lemme start with some basics, I guess. Do higher resolution images fed into the training produce images with better detail/fewer artifacts?

[astral goblet]

oh geeze. first day back so i can ride the sd3 preview, and allready the first link shown is a full on furry porn model

doesn't nsfw shit belong on other servers?

[forest trout]

Well you are early. SD3 isn't out yet.

[astral goblet]

i'm mostly okay with furries. they're doing their thing and mostly not hurting people. but wow . WOW. they're the worst kind of exhibitionists

[astral goblet]

wait list asks for the discord user so i assume you've gotta be here

[brave vigil]

😐 I just wanted to provide the specs of the model. I gave y'all a disclaimer

[forest trout]

Ah, forgot about that. Nevermind then.

[astral goblet]

i didn't even scroll. it was right there

[brave vigil]

Shit - sorry

My bad

[fervent thunder]

I have the link, you can delete I suppose.

[forest trout]

I mean civitai isn't exactly great about censoring their outputs.

[astral goblet]

if you're gonna go buck under your bathrobe in a room full of people, keep that shit tight

[brave vigil]

Sorry about that all. That's my bad

[fervent thunder]

One message removed from a suspended account.

[brave vigil]

I'm still drinking my first cup of coffee -_-

[fervent thunder]

@brave vigil Okay, back onto the tecnical side of things - that model is SD 1.5.x-based, which means it has 512x512 native resolution, which means you'll want only 512x512 training data.

[astral goblet]

settings are highly dependent on what your dataset and captioning approach is. what your goals are for the model. etc.

[sudden ruin]

Where?

[astral goblet]

not on yet. i'm just here waiting

you're a mod you know!

[fervent thunder]

One message removed from a suspended account.

[forest trout]

He's asking where the NSFW is.

[fervent thunder]

One message removed from a suspended account.

[sudden ruin]

Sounded like there was a link already here to me

[forest trout]

He's talking about the civitai link.

[astral goblet]

lol oh. the civit link lol. i wasn't even that worried. just non chalantly surprised. didn't mean to rile mods babout it. he deleted i think

[brave vigil]

I'm glad I decided to ask for advice instead of just working it. I've got a number of questions based on what you're all saying.

Firstly - when I create an image using my current checkpoint at 512x512, the output is pretty bad. The concept is there, but there just aren't enough pixels to get details. Given the chance I prefer to have 1024x1024. Is there some way I can scale down the actually decent bigger images without losing quality?

[astral goblet]

often when people say "nsfw" i expect some boobies but not raging erections. just my gender bias outrage happening maybe

[sudden ruin]

Feel free to report if you come across something similar again

[astral goblet]

it was pretty tame but thanks for the heads up.

[brave vigil]

Also - someone once told me that when training a LoRA, I should do so on the base model of whatever checkpoint I'm using. My checkpoint is built off SD 1.5, so when I boot up Koyha and give it a model to use in training, should it be the base model or the checkpoint?

[fervent thunder]

See, I joined this chat, and mentioned using an image of a Mexican boy holding a Crucifix as a tool against sketchy NSFW content, someone rolled his eyes at me for doing so, but it turned out to be needed not five minutes henceforth. I posted it in general_with_images if anyone needs it.

[astral goblet]

its in #🌠｜show-and-tell

[fervent thunder]

Base model for a "clean slate". It's pretty easy to corrupt a checkpoint beyond recovery by making a silly mistake in training (like, not having the dimensions of your images be divisible by 8), I've done it many times, so I try to start from a clean slate when possible.

[brave vigil]

And a third question. If I'm having trouble with detail/resolution, should I consider moving to a different checkpoint or even to SDXL (not sure what makes SDXL different, tbh)

[astral goblet]

Oh. Its mocking autistic people. huh i'll report shit like that. I have friends with autistic siblings and people mocking them for who they can't change being is one of the worst parts of their life.

[oblique edge]

you're in the wrong discord

[fervent thunder]

"Autism" is essentially a rebranding of mental retardation. Look at the statics.

[astral goblet]

This is where the sd 3 preview will be? i'll leave again after its over

[brave vigil]

WRONG

This is wrong in so many dimensions

[fervent thunder]

Also, genome-wide association studies show essentially complete overlaps between retardation and "autism".

[astral goblet]

lot of hateful language around autism today. @slender fiber i think this stuff is worse than any nsfw content by a moon shot

[brave vigil]

Wrong by definition. Retardation refers to brain damage or malfunction. An autistic brain works totally fine, just very differently

[oblique edge]

There will be no plublic preview, just a release

you have to run it locally

[astral goblet]

https://stability.ai/news/stable-diffusion-3

there'll be an invited prview, suspected to run as a secret channel on this server

[brave vigil]

Also - I realize why I was so nonchelant about sending that link before. I'm in another server specifically for furry SD artists. The profile icon for it is also purple and abstract, and I genuinely thought I was posting in a furry oriented group. Like I said, first cup of coffee

[oblique edge]

While the model is not yet broadly available, today, we are opening the waitlist for an early preview. This preview phase, as with previous models, is crucial for gathering insights to improve its performance and safety ahead of an open release. You can sign up to join the waitlist here.

2nd paragraph

gotta read it

[fervent thunder]

@brave vigil Evidence part 1: Look at statistical rates for retardation and autism among school-aged children in the United States (lots of good data). Autism rates "skyrocketed" exactly when retardation rates began to drop. Also, this happened first on the East Coast, in New England (the most WASP and wealthy part of America), then spread to the West Coast (the second-wealthiest area), and then the interior. In all three zones, the pattern was the same: diagnoses of retardation dropped as those of autism rose, such that the sum total of both rates stayed about the same.

[astral goblet]

ah yes. cross servering. dangerous times. you're all good bud. i was joshing you mostly. 😉 furries are all good with me

[brave vigil]

As John Oliver has said many times

Furries are harmless. And hilarious

[astral goblet]

right. i did. now go read where i said there was going to be a public release? i'll wait

[oblique edge]

@skywalker if he's really from germany he's going to have a fundamentally different approach to mental handicaps with regards to medicine

[oblique edge]

I'm not waiting I'll just block you because I no longer wish to provide assistance to you \

[brave vigil]

#🌶｜off-topic message

[astral goblet]

John Oliver was a great furry ambassador, but what got me to really stop caring about them was Violent J from the Insane Clown Posse. Those are another bunch of cosplayers that have fun and hurt no one and his daughter is a furry now, so he's an ambassador.

[astral goblet]

weren't really helping in the first place big gunner

[oblique edge]

touche'

none the less you seem opinionated and very eager to jump down other people's throats with regards to touchy social subjects. I wish you well.

[fervent thunder]

I have an amazing ICP story (the band, not the crypto). It was my first contact with "the Internet mob" that is so common now.

[oblique edge]

I used to listen to ICP in high school, if you're still listening to ICP in your 20/30's you need a reality check.

[fervent thunder]

Heh, it's all good, so did I.

[astral goblet]

used to watch em on wcw and wwf. fun times. wacky guys

[brave vigil]

I'm going to jump into #🔧｜finetune to ask my technical questions

[oblique edge]

They had some good album art... The Great Malinko, Riddlebox. It'd be interested to train a model JUST to make ICP album covers. Take a prompt and turn it into one of those heads from their albums

[forest trout]

Probably could just look up the original artist and find more of their work.

It's like how so many of the classic rock musicians all used the same artist: Journey, Boston, ELO, Foreigner, Asia, etc.

Found them: Shusei Nagaoka

[astral goblet]

Love those boston album covers!

[low moon]

Is there a plugin for Stable Diffusion that lets you apply different clothes to a character? Not IP Adaptor and inpainting, that doesn't work well. I mean like Reactor for faces. You upload images of clothing items and let the AI do the rest. Feels like something that should exist by now. Like Outfit Anyone.

[lone hawk]

SD3, tomorrow something is gonna happen: https://twitter.com/EMostaque/status/1764742280865284128

[jovial wraith]

oooohhhh!!

https://x.com/EMostaque/status/1764743952526119380?s=20 "Tomorrow or Wednesday"

[charred mesa]

LET SGO

ÉLETSTSGBJKSdg

[winter pike]

omg finally

[fervent thunder]

Once upon a time, I read an article about how Juggalo make-up (cosmetics) was effective in defeating computer vision (sort of a precursor to more modern AI facial recognition). Basically, a group of researchers at Stanford University decided to paint people in Juggalo make-up to see if it defeated facial recognition, and it did.

So, learning this, and knowing about the famous ICP "magnets...how do they work?" line, I made a meme based on the "weak SpongeBob/strong SpongeBob" template, where the weak version was "Juggalo knowledge of magnets" and the strong version was "Juggalo knowledge of facial recognition algorithms".

I thought this was a harmless joke. I was wrong. I had angered the ICP army. Suddenly all my socials and inboxes were full of death threats for insulting ICP and Juggalos. I explained that the meme was actually sort of a compliment - like, that Juggalos had the foreknowledge to develop cosmetic face-paint to defeat AI facial recognition. But, the Juggalos were having none of it. My first experience of an Internet mob. I'll never forget it. Got Doxxed and legit lived in fear of a Juggalo attack for like a month.

[mortal delta]

am i allowed to ask about hobbies recommendations here or would that be better for off-topic?

[jovial wraith]

i think off-topic would be better

[mortal delta]

just wanted to make sure...

[fervent thunder]

Correct, this chat is intented for discussion of how, and to what extent, Juggalo make-up interferes with facial recognition systems.

[broken cave]

lol

[forest trout]

This inspired me to set the newest #🔆｜dailies to music album covers inspired by that style.

[steep timber]

how can i convert a repository to a .safetensors file? Like, an model i want to use does not have the final .safetensors and I would like to make one

[astral goblet]

you weren't a victim. you're a dude who punches down. you were likely doing that intentionally to the fam you angered (they really do consider themselves a family) and laughing about it. "Death Threats" were almost certainly instigated. You're a giant cliche . "Reee" gave it away.

[fervent thunder]

I would never claim that all (or even most) Juggalos are troublesome; in fact I've rarely met any. Perhaps the offenders were agents provacateux, engaged in a false flag anti-Juggalo operation. I wish them all the best - it was more a funny moment for me, especially in retrospect, and I wish Juggalos success in defeating computer vision. If SKYNET exterminates all humans except Juggalos (assuming it is unable to recognize the latter on account of their make-up), and my bones are condemned to the dust-bin of history as the Age of Men ends and the Age of the Juggalo begins, that would serve me right.

And, to be fair, I do know now that the "magnets...how do they work?" line wasn't intended to imply that ICP thinks that magnets are magic; it was really just an expression of the idea that so much of the natural world is amazing (magnets included), and that the "magnets..." meme is taken out of context.

[oblique edge]

Lol I see the convo has drifted to this discord

Ahhh ICP

I knew their music was trash, I just liked the Dark Carnival from a lore perspective

[astral goblet]

holy shit. wasn't expecting this to show up. https://www.reddit.com/r/StableDiffusion/comments/1b6ivqg/coherent_multigpu_inference_has_arrived/ NVLINK diffusion code

[oblique edge]

I'm thinking of using it as the basis for my AI dungeon master

At least the lore part of its MoE

[fervent thunder]

Very impressive RE: Multi-GPU. Sadly I am likewise unable to test it 😭

[astral goblet]

I enjoyed them from the sidelines. Lots of juggs at my school but i thought their music was lame haha. couple good ones at least. To me they were always the clowns from WWF

[fervent thunder]

Would be cool to have a pair of 4090s, but...I don't, haha.

[astral goblet]

4090s don't have nvlink

[oblique edge]

I have a 3080.... working on talking my fried into letting me offload compute to his 4000 series as needed

[shell tendon]

yehawwww i'll be getting a pair of 3090s someday then

[astral goblet]

i got a 4080. is good.

[fervent thunder]

Is it like, a workstation-only thing? Like the thing about Martin Shrikelli (sp?), the "pharma bro", in the clip where he's hitting up the ghetto to buy some bootleg H100s?

[steel willow]

General question, I just got a 4080 super and only downloaded drivers... I'm using ForgeUI, do I need to also download CUDA?

[oblique edge]

Yes I think

Webui.bat should handle it?

[astral goblet]

nvlink is on ampere cards. they left it off ada cards though

[fervent thunder]

I think so too, but could be better to let StableDiffusion UI (like Automatic1111 or whatever you use) try downloading it, it will get the right version, I've had a million issues with wrong versions.

[oblique edge]

I dunno I set up my shit so wrong at first so don't listen to me lol

What he said

[astral goblet]

it helps to install the cuda sdk imo. i got 11 and 12 installed

[steel willow]

Hm how would I know if it has?

[fervent thunder]

@oblique edge Same, set up my shit wrong a million times. And what I said was what you said, let Webui do it.

[steel willow]

Atm I get Device: cuda:0 NVIDIA GeForce RTX 4080 SUPER : native

Which was weird

[fervent thunder]

I think "cuda:0" may just mean that the card supports CUDA (i.e., you're in GPU mode, not CPU mode, which is good) and that it's device zero (the first number).

[astral goblet]

use a package manager to install stable diffusion. easier than dealing wiht the terminal commands yourself. stability matrix works good.

[steel willow]

Don't get me wrong though, I'm still generating fast but just don't know if I actually have CUDA XD

[steel willow]

That's what I did with Forge

[steel willow]

Ah that could make sense

[astral goblet]

scroll to the bottom of your webui and look at the pytorch version

[oblique edge]

As long as it's not bottleneck'ing anywhere and console shows it using VRAM your probably fine

[steel willow]

top?

Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug 1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]

I got this

[astral goblet]

very bottom line

of the webui

[steel willow]

Idk why it says AMD (must be CPU)

Oh bottom?

[astral goblet]

all 64bit cpus , intel or amd, are amd64 architecture

[fervent thunder]

@steel willow AMD64 is the instruction set (x86)

[oblique edge]

Yeah cpu

[steel willow]

I don't see pytorch version 😮

[oblique edge]

It's not the newest pytorch I can vouch for that lol

[astral goblet]

API  •  Github  •  Gradio  •  Startup profile  •  Reload UI

version: f0.0.17v1.8.0rc-latest-273-gb9705c58  •  python: 3.10.9  •  **torch: 2.1.2+cu121 ** •  xformers: 0.0.23.post1  •  gradio: 3.41.2  •  checkpoint: 912c9dc74f

[steel willow]

OH

ON THE WEBUI

[oblique edge]

It should dl the proper version

[steel willow]

I thought you meant bat

[oblique edge]

Yes

[steel willow]

torch: 2.1.2+cu121

Thanks, all good

[astral goblet]

you got the coods

[oblique edge]

Looks good

[steel willow]

Thanks for the help!

[oblique edge]

Np

[oblique edge]

How do I avoid multiple instances of something in my output? For example every time I give it a simple prompt like "a toy soldier" it keeps putting 3 in each image

I've tried using sole, singular, individual... even.put things like many, multiple, etc in the negative prompt

Still having issues

I know this is inherent to sd2? should be less of an issue with3? Or am I missing something?

I'm also not well versed in sampling methods so maybe I'm doing something wrong there

Also need to brush up on my art terminology, has anyone found a good index of terminology I could reference?

Surely an "art dictionary" as it relates to generative AI already exists

Or at least I'd be willing to pay for one

[patent scroll]

Does anyone have an idea how to make A1111 use the shared memory as well?
I don't need it for normal generation but it appears 16GB VRAM is not enough for 4x hiresfix and it could use half (16GB) of my RAM there.
Even if the upscaling would be slow, at least it would work and not throw OutOfMemory

[hallow sable]

auto sucks

[astral goblet]

nvidia drivers provide this option by default. it has to be at the driver level. i assume you have amd. don't hold your breath for them to get there

[astral goblet]

skill issue

[hallow sable]

not skill issue, auto is for low skills

[patent scroll]

Yes AMD, and yes, kernel level. windows is able to do it, potentially via the system API, therefore it's possible.

[tidal yarrow]

so it isn't letting me run stable diffusion from the .bat file but it is letting me run it on the "launch" file. does this affect anything or nah

[pine fiber]

anyopne got sd3 yet

[pseudo bough]

hey guys whats currently the best way to colorize/upscale old videos

so i found deoldify on forge

looks pretty good now i gotta find the bets video upscaler out there

[pearl ocean]

I hear for, welcome our A.I overloads

[night cipher]

Hey Guys,👋 as a part of my senior Design Project, I have created my own Image Generation Model and have deployed it on my platform, if anyone here wants to test it out please do : https://thetrazo.com/dashboard?show=1

[pine fiber]

topaz since i last checked, but i bet theres better now

[shell tendon]

workin on a custom face swap workflow... better than any node i've used at the moment (requires a lora, but a lot better than a naive lora swap)

[patent scroll]

Benefits of Forge over A1111 at 16GB VRAM?

[nova zodiac]

A4000 gave me an extra 6it/s on forge

[patent scroll]

OK, how about benefits except the speed?

[nova zodiac]

The supermerger extension didnt work for me, but the inbuilt freeu and latentmodifier are very nice

[tidal yarrow]

why doesnt stable diffusion say i have xformers even though when i run "pip install xformers==0.0.24", it says that all the requirements are already satisfied

[fervent thunder]

I believe you need to set some sort of flag when starting WebUI

[tidal yarrow]

how so?

[nova zodiac]

Do you have —xformers in the webui-user.bat file?

[fervent thunder]

What ResidentChefAU said.

[tidal yarrow]

set COMMANDLINE_ARGS=—xformers

is that what it supposed to look like

[nova zodiac]

Thats RCNZ to you 😉

[tidal yarrow]

or is there a space

[fervent thunder]

I usually create a shortcut with the custom settings (actually I have several shortcuts, so I can launch in different modes).

Your shortcut would be like webui.bat --xformers

[brave vigil]

Question for y'all

[astral goblet]

its a driver problem. windows can't do it on it's own. it would rely on the driver.

amd windows driver are grade b

[brave vigil]

I want to make 2d illustrations, right? And honestly, I kinda want the quality (but not the resolution) to be low

[tidal yarrow]

wait but what should the notepad look like'

[brave vigil]

I want it to look like something an intermediate artist would make - not something polished

[astral goblet]

the unet patcher and the extensions that will inevitably be written for that

[fervent thunder]

@Luke By "low-quality", do you mean low-poly?

[brave vigil]

No...

More like...

[fervent thunder]

Ah. I would just cut the number of iterations then...I have always been extremely successful in creating low-quality images, I don't mean to brag, but it's true.

I would also try different samplers, DDIM instead of Euler-A, for example. DDIM isn't necessarily "worse" but it's less predictable, which could be desirable.

[brave vigil]

Hmmmm

To be clear though, I think I'm maybe talking a bit more about stylization

On one end, you've got doodles

On the other, you've got highly photorealistic digital paintings

In between you've got your average art style, with exaggerated proportions, unique line work, etc

And that's what I'm trying to get (without borrowing from existing artists)

I've tried invoking prompts like "comic," "disney," "exaggerated proportions," or "cartoony"

With some success

But what I could really use is a primer giving the names of specific cartoon/comic artsyles

[astral goblet]

if you're using something like juggernaut or other photoreal models, gonna be hard to prompt them to toon styles.

[brave vigil]

Just normal SD 1.5 on Auto1111

[astral goblet]

"Disney" is the style really since they basically pioneered western animation. Other studios work too like style of ghibli, but each of their movies is so unique that it helps to go even further and "style of a movie name" too. Artist names help a lot. Bluth was big in the early days of disney and left to do things like Titan A.E.

[fervent thunder]

Do you mean like caricature? That's a distinct (and, I think, pretty cool) form of art, created by Italians in New York to sell to tourists. I don't mean this as a knock against it; the fact that it's popular can be evidence that it's good. There's also overlap between caricature and propaganda, very interesting.

[brave vigil]

Titan A.E. was AMAZING by the way

Maybe not so much for the story, but from a production level

[fervent thunder]

Woah, you know what would be cool? A CupHead-trained model.

[brave vigil]

https://tenor.com/view/perfection-perfect-michael-fassbender-magneto-x-men-first-class-gif-2260686001358167676

[brave vigil]

This is an interesting idea

[astral goblet]

why use the base 1.5 model? it kind of sucked imo. Runway ML rushed it out to piss off Emad for some reason. All the drama was contained behind closed doors but they got fired for dumping it on hugging face

[brave vigil]

There are a number of "niche" or "hyperstylized" art styles which, if turned down to a simmer and paired with regular disney animation, might help spice things up

Graffiti style

[fervent thunder]

Oooh, that's another good one.

[brave vigil]

You're using words I don't understand

[astral goblet]

yeah i agree. The story wasn't badbut it sure was campy. SO MUCH HIGH QUALITY PRODUCTION. Bluth knows his shit

[brave vigil]

I've been doing this about 2 weeks. I think I'm doing okay, but all I can do is use the resources I'd found to work so far

[brave vigil]

Daddy issues. Daddy issues everywhere

buzz lightyear

[astral goblet]

you should look into community refinements. Theres been a lot of work done

Fun Fact. Don Bluth is working on getting a Dragon's Lair movie made

[brave vigil]

#🏞｜general-with-images message

[fervent thunder]

"Base 1.5" is a 512x512-native model. It's not the most up-to-date, but probably has the most support in terms of LoRAs (though that will change). FlowWolf is correct, SDXL is technically superior in every way.

[brave vigil]

What I'm trying for is a delicate mix of western cartoon and realistic-ish

[fervent thunder]

This is going to sound extremely stupid, but...sometimes the easiest solution is to hide hands (hands-in-pockets, etc.).

[astral goblet]

not just sdxl. All the community refinements like anything v5 or dreamshaper or protogen

[brave vigil]

I like the exaggerated artwork from comics and cartoons, but also the realistic proportions and relatablity of realistic work

[fervent thunder]

Sometimes the easiest approach to a problem is not to solve it, but to cover it up and pretend it doesn't exist. If you have a problem you need not solved, I can help not solve it.

[astral goblet]

Artists have been phoning in hands for centuries!

just phone it in

[brave vigil]

Oh man

This gives me feels

[astral goblet]

Nobody cares that a dragons lair movie staring ryan reynolds is on it's way ?

[brave vigil]

The front of my brain knows that in 99% of situations this is totally correct. My background is in software development though, and this is something that's beaten out of us pretty thoroughly

[tidal yarrow]

how do i keep it so my settings save? its annoying having to re-set it up everytime i relaunch sd

[astral goblet]

hethens

[brave vigil]

Errors should never pass silently unless explicitly silenced - which is I suppose what you're talking about. As a general rule, I prefer the philosophy that "if the problem is to hard to solve, change the problem"

[astral goblet]

if (error) continue;

[fervent thunder]

I believe that not-solving problems is key to success in the AI industry, especially as pertains to moderation. Concerned that your AI produces "objectionable" text when asked about certain issues? Just get between the actual AI and the user, and set it so that, if certain tokens are detected, the AI will never even activate, and instead a static output of "I'm sorry, but X is a sensitive issue. One should always be respectful when discussing X..." is returned.

The same thing with Gemini. Not enough diversity in source data (arguably, an actual and significant problem)? Instead of changing the training data set, just "get between" the user and AI, and inject certain terms into the prompt in order to "diversify" the output, regardless of what the prompt is and regardless of whether this makes sense logically. Then, when people complain about this, just disable the ability to generate images of people entirely.

Non-solution of problems is a key approach at the highest levels in software development, finance, and government.

[astral goblet]

@brave vigil https://civitai.com/ peruse this website. There are 100s of community refines of sd 1.5 and sdxl . just be careful of all the community nsfw obsession

[brave vigil]

Ahem 😓

My older sister always says - and I believe this too

Perversion is its own reward

[astral goblet]

feels culty to me

"leader is always love"

[fervent thunder]

Indeed; intercourse is one of the few activities that cannot be engaged in ironically. Almost everything else can be done ironically.

[brave vigil]

Ohhhhhhhhhhh that is good

[astral goblet]

basic biological actions. shitting too

hard to sleep shit fuck eat ironically

[brave vigil]

I was raised "anti catholic." My parents were what you might call refugees from devout catholic families. They raised us to believe that hedonism is NOT sin, rather, it is the wine and olive oil which give life its true flavor

So go ahead - enjoy your shoe fetish! Squeeze every bit of flavor you can from the lemon that is life

[fervent thunder]

I think it's important to remember that the most important use of irony, is inspiring online arguments about whether something is actually "ironic" or not.

And the fact that 90%+ of instances of irony are related to Internet arguments about whether something was actually "ironic", or just "unexpected" or a "twist of fate", is itself ironic.

[astral goblet]

Everything in moderation. Being catholic and then going full hedonistic might've been an over correction. I see the same thing happen to JW kids. They grow up and hate being "proper" so they swing the other direction hard.

Enjoying yourself is fine. We're in the 21st century age of abundance. We aren't quite at 31st century hedonism bot levels though.

[astral goblet]

disagree. Singing Morsette is the highest calling

https://www.youtube.com/watch?v=Jne9t8sHpUc

[brave vigil]

I'm exaggerating a bit. Or at least, I'm only explaining the part about the hedonism. Common sense was another important theme

[astral goblet]

lol hedonism bot is an exageration too. i love the guy

[fervent thunder]

So, "reading the room", it's my sense that there aren't many Apostolic Anabaptists (Amish/Mennonites) in this chat...

No, I mean, I am Amish-pilled both because I think they foresaw some of the hazards of technology (to be fair, so did American Indians - maybe the Lakota were trying to prevent themselves from being "trained" when they thought that "cameras stole the soul" in the 1820s), and also, my guess is that, living in a tight-knit community and embracing creampie maximalism and large families is an optimal approach to overall life satisfaction, as well as continuation and indeed growth of one's "tribe" or group, howeverso defined.

I won't go further into the demographic considerations of creampie maximalism as a cultural strategy, out of (legitimate) fear of being banned by someone whom'st've's'd'nt realized that any technology which has the effect of reducing the number of living persons of a given ethnic group ought to be viewed with the same suspicion as a file which ends in the extension ROM~TILDE/PUB/CIA-BIN/ETC/INIT.DLL?FILE=_AUTOEXEC.BAT.MY%20OSX%20DOCUMENTS-INSTALL.EXE.RAR.INI.TAR.DOÇX.PHPHPHP.XHTML.TML.XTL.TXXT.0DAY.HACK.ERS(1995)_BLURAY_CAM-XVID.EXE.TAR.[SCR].LISP.MSI.LNK.ZDA.GNN.WRBT.OBJ.O.H.SWF.DPKG.APP.ZIP.TAR.TAR.CO.GZ.A.OUT.EXE

(by which I mean "sus")

[elder nimbus]

Hey guys, do you know when the AI image generator will start working again?

[fervent thunder]

Are you referring to a generator as part of this Discord "server"? Or some service by SD delivered through their site? Or their API?

[elder nimbus]

yes that of the discord server

[fervent thunder]

Hmm, can yyou provide an example of a prompt you would like to test?

[winter pike]

allegedly more news tomorrow or wednesday

[elder nimbus]

@winter pike Ok thanks for the information 👍

[shut jasper]

#💬｜general-chat

[crimson beacon]

Any news on SD3 release or at least a beta? As of right now I'm okay with seeing lots of outputs of the model

Super curious

[hallow sable]

comes out when it comes out

[oblique edge]

Assume unlimited compute. Assume I have a 135 IQ for output efficiency. Here you will operate as my stable diffusion guide. While I study, keep in mind that I have very basic knowledge of art terminology. Therefore you may use as much advanced terminology as needed regarding artistic styles, terms, equipment etc., as long as you define it. I will use this to explore the various options as I teach myself the technology. We'll be focusing on fine-tuning detailed settings available to the various sampling methods. Assume I want to output HD images only. Lets start by walking me through making my first image.

This has proven helpful

[pine fiber]

Who said that

[oblique edge]

I use Gemini
Leave out assume unlimited compute if you don't have a higher end card
or just insert "I have a [whatever card] with Stable Diffusion [whatever version] installed locally" in between the first and third sentence
This functions well for any of this new tech you want to learn about
It's kind of plug and play
Keep in mind Gemini only updates like every 3 months but if you're on that much of the bleeding edge of this stuff you're probably not hanging around here lol

[winter pike]

https://twitter.com/EMostaque/status/1764743952526119380

[pine fiber]

Thanks lol

[candid jacinth]

d

[astral goblet]

Grade schhool used to be set up to teach people to self learn. Seems now a days all that happens is people come out of grade school hoping that other people will teach them everything. The way i heard it, you're here to "learn how to learn".

Theres so many learning opportunities on the internet. I dont understand how people still need a personal guide and expect it for free. Soon chat bots based on hallucinating models are going to be teaching everyone cause they don't know how to self learn. So we're going to end up with some really fucked up hallucinated metas

like "Its what plants crave"

[shell tendon]

agreed on the latter... though i'm not sure grade school ever set anyone up to be an independent thinker

[fervent thunder]

Yep. Back in the day, we thought that the limiting factor in human progress was information availability. Turns out, it's not that.

[upbeat marsh]

how to use stable?

[cinder dune]

is there a way to merge sdxl models with 1.5 ones?

[shell tendon]

no

[brave vigil]

CivitAI is without a doubt an important resource

But as a gay man with perhaps a bit or objectivity

Holy tits, there's a lot of tits on this site

They should have called it TitsAI

[pine fiber]

the more the merrier

[brave vigil]

XD

Don't get me wrong

I'm pro tit

Release the Tit!

But there's a limit. Like salt

Quick aside. How robust is SD's NL understanding?

It certainly doesn't have the juice of, say, ChatGTP (though I don't see why)

You're joking, right?

https://tenor.com/view/furry-owo-furry-owo-gif-20975753

I came for the fur

XD Yeah, I'm that guy

https://tenor.com/view/pickle-dill-with-it-gif-8229752

Maybe you can help me understand

Why is it

That with every model except for one that I've ever used, no matter what prompt I give (even ones copied word for word from other posts on civit), all I get is garbage

[karmic brook]

can you guys take tht kind of discussion to dm if nsfw

[brave vigil]

Not NSFW. Just dicussing the fandom, and how the tech can be used in a general sense

[astral goblet]

cant really recommend people civit without warning them

stable should make a host

we can talk about the safe host then

"Hugging Face" is not one either, since the bar is civit. For the community purposes. Capiche?

i can't talk about it sorry

[brave vigil]

Every other model I've ever used, and I can give the same prompts and get nothing but garbage

Lemme rephrase in more community acceptable terms them

Anyway, I use this one model - X. Every other model I've ever used takes the same prompts and gives me nothing back. What's the deal?

[astral goblet]

cfg too high or low. steps too high or low.

[brave vigil]

Or maybe just bad prompts? Not speaking the right language?

[astral goblet]

any prompt should make an image thats not just noise and garbage crap

[drifting flame]

couldnt agree more

[slow merlin]

hi, not sure if it's right channel but we paid $20 for commercial membership (twice!) but don't see the same reflecting in our account. It's still the basic 25 credits account. Stability have specified their support emails but getting no response at all (been 3-4 days). Would anyone have an idea or faced something similar? Is there a way to subscribe from within your account (membership payment was done via https://stability.ai/professional)

[verbal delta]

I am not sure whether this goes under off topic, but I am wondering whether there is a way to customize your own voice without needing to use someone else's voice? I am looking for voice modifies with settings such as for timbire, pitch, masculinity to femininity, etc.

[pearl ocean]

S o r a

[fervent thunder]

@admin

brother

I'm sure you "didn't notice"

but this baguette

has veins.

[fervent thunder]

[hallow sable]

trump commemorative xbox mountain dew gamestop collectors edition coin

XL with pineapple

why no make image???

[halcyon granite]

I'm following an auto1111 tutorial using tiled vae + high res fix to output a large 4k image. I'm using a 3090 with the same paramters that they used with a 3060, problem is i left it running over night and the image was stuck generating for 330mins and never complete. I'm wondering what I can do since my spec should be more capable, maybe one of my command lines is making performance slower? "set COMMANDLINE_ARGS= --autolaunch --xformers --no-half --disable-nan-check"

[warm junco]

Yes your slowing yourself down with using --no-half and --disable-nan-check is also bad
You should add --no-half-vae
Then Tiled VAE should work

For more help feel free to post in #🤝｜tech-support

[pulsar marsh]

ay what exactly do you use ControlNet for

[grizzled palm]

Largest size SD3 8B on the 4090!!! In early, unoptimized inference tests on consumer hardware our largest SD3 model with 8B parameters fits into the 24GB VRAM of a RTX 4090 and takes 34 seconds to generate an image of resolution 1024x1024 when using 50 sampling steps.
Sure, that's longer than Cascade takes for 4K, but we're talking about 8B parameters.

[ivory dagger]

Hi

can anybody please help me fine tune my lora/dreambooth model?

I'm using abstract art like my profile picture so I think it's a bit tricky for the AI to figure out

any help appreciated!

[grizzled palm]

Since SDXL is trained to maximize an aesthetic score, it has a really hard time learning to make ugly art.
(LOL sorry I'm just kidding. I have no idea how to train loras.)

[ivory dagger]

noo don't call it ugly

wym

[grizzled palm]

It was just a joke I promise. I can't help.

[still glacier]

try asking in #🔧｜finetune

[ivory dagger]

i actually believed u for a second about the aesthetic score

Lol

or is that part true

[ivory dagger]

ok i will thanks

[grizzled palm]

The aesthetic score part is true. But it can definitely learn amateur/ugly anime in a second. Training isn't a problem.

[ivory dagger]

Yeah, but I'm new to all of this

I mean I'm excited it's just that this is probably harder to get used to the whole fine tuning thing

[still glacier]

it s definitely harder to finetune than to generate

[ivory dagger]

I mean yeah

[still glacier]

totally ok to get lost at first

[ivory dagger]

but I meant like my goal would be easier if i was just inputing my face or something

this art style is kinda fucked

if you look at my training data it's a bit too harsh i feel like

[still glacier]

I have'nt watched it yet but you can t go wrong with aitrepeneur https://www.youtube.com/watch?v=1BCYdd9r1To

[ivory dagger]

yeah i found his channel yesterday

thanks

i'll skim through it can't listen to audio rn

but hmm how many epochs should i train with 287 images and 15 steps

can you give me a rough number rn?

i'm trying my larger image set again

[crisp shadow]

d

[fair cave]

Hey guys! If you're looking for a different image generation UI, check out: https://www.practal.ai/

[molten slate]

Hi, does some one know how to get a API key for image inpaint? It is for a schoolproject 🙂

[flint lark]

When will the bots be back

[ocean bone]

Question for the Stability AI team, if you are reading this:

Now that the SD3 paper has been published, can you provide some more guidance on the VRAM requirements for both inference and LoRA training of the 8b parameters model?

I saw a table toward the end of the paper, but I'd like to double-check. Thank you!

[fair cave]

You can check this out: https://platform.stability.ai/docs/features/inpainting#Python or use https://www.practal.ai/

[molten slate]

Yes but can't click on inpainting?

[nova zodiac]

They said the full fp32 version will work on a 4090 24gb vram so fp16 should work on 12gb cards

[ocean bone]

this is for the 8b model, correct? Thank you.

[nova zodiac]

Correct

Which makes sense as you can get sdxl on a 6gb card with some room to spare and thats 3.3B

Haha no, but if you have that much spare $$$ burning a hole in your wallet feel free to send a 3090 these ways!!!

😛

[agile nacelle]

Are you able to XYZ Plot Mask blur settings in In painting ?

[flat sparrow]

I’m still new to SD. What is this fp32/fp16 thing ? If I download a checkpoint model, how do I know which one is it and how will it influence my generations?

Another question is regarding sdxl. There seems to be several versions. If I download a Lora for the base sdxl from civitai, will they work for lightning or turbo versions?

[nova zodiac]

Fp = floating point - the 32 and the 16 refer to how many bits (ie how accurate) is being used to store each number. Will influence but its a game of wheres wally to find the change usually.

[nova zodiac]

Turbo no, lightning probably

[fervent thunder]

Lots of barley grains bouncing against a gradient dark brown background

[sharp meteor]

Hi, is there a way to set up Focus to open in a specific browser without having to have that browser as default in windows? How, if it's possible?

[jovial wraith]

i don't think that's possible

afaik it generates a url in the form of an ip which windows recognises and then opens with the default browser. i could be wrong tho

[sharp meteor]

interesting, I've found this but I dunno where to put that line

--disable-in-browser

looks like something I'd put in the webui-user batchfile if it were A1111 but can't find anything like that in Focus

[jovial wraith]

i'm afraid i can't help you with that then, as i don't use fooocus

[sharp meteor]

thanks for trying tho

[sudden ruin]

Thats pretty long, damn

Got a source?

[sudden ruin]

Sit tight and wait for their answer, they probably have a lot on their plate right now. If you dont get an answer within the next few days, feel free to open a #1010934719455707218 ticket and ill try to elevate your issue!

[grizzled palm]

Check announcements they released the research paper last night. (Source.)
Time should come down with xformers. If a lightning version is trained, it should drop to 3 seconds.

[thorny sedge]

hi

[jovial wraith]

heya!

[worthy bone]

Hi Guys, are you saying that today will be the lucky day where they will let us access the preview of sd3?

[full lark]

I hope they do 🥺

[worthy bone]

yeah 😦

[jovial wraith]

according to what Emad said, either today or tomorrow. source: https://x.com/EMostaque/status/1764743952526119380?s=20

[worthy bone]

Let's hope so 🙂

[jovial wraith]

but needless to say, that will be a private discord with a bot running SD3. if you let people run it locally they can just go around sharing the model

[worthy bone]

Do you think it will be the discord bot?

I mean if they will do a preview with the bot

[jovial wraith]

not necessarily the one in this server, but rather a private bot for people who joined and got accepted into the preview

that's all just speculation tho

[worthy bone]

I understand.
On Twitter I saw that a person had been accepted

[jovial wraith]

wait really? do you have a link?

[worthy bone]

yeah

I don't know if I can share the posts here on the group but just type #sd3 and go to latest

[jovial wraith]

i see. i'm skeptical tho, since the image quality isn't anywhere near what others have shared

[worthy bone]

I agree with you, in fact it seems very strange to me

Oh maybe those images refer to a past technology

[jovial wraith]

that's also a possibility

man, i hope the requirements for SD3 aren't too high. at least for the smaller model. i really don't have money for a new gpu right now

[worthy bone]

I understand you perfectly

I have a 12GB 3060

[jovial wraith]

i have a 3gb 1060 xD

[worthy bone]

Maybe you should be able to handle the 800m one, I hope 😄

[jovial wraith]

that would be the ideal outcome. anything higher than that isn't going to be it

[worthy bone]

We can only wait and understand. In fact, 3GB is very little nowadays

[trim magnet]

no way to know unless you work there 🤷‍♀️

[jovial wraith]

very true. if it weren't for the fact i wasn't into ai when i bought it and it was a great deal, i would've gone for something better

[full lark]

Are you sure that the image was done in sd3? The person who post it said "VQGAN + CLIP" in the end

🤔

[trim magnet]

ill direct this to the ml engineers team at sai so they can send you a memo next time they come up with an improved training script

[worthy bone]

Until last year I had a GTX 1080 8gb, then I switched to the 12GB 3060

[full lark]

It depends on whether they buy the hardware or just rent it

[jovial wraith]

i almost bought a 6700xt, boy am i glad i didn't

[worthy bone]

You did well 😄

[trim magnet]

now imagine if amd improves rocm on windows

[jovial wraith]

that would be great, i don't think it's going to happen anytime soon tho

[near silo]

and if we used 100% of the entire worlds wafer capacity, it would still take close to 2 full years

[trim magnet]

dam emad exposed

[near silo]

I have a feeling they are measuring in a different way that isn't being explained right

My guess is they are referring to their total systems FLOP potential in 1 second multiplied by the number of seconds trained, then flattened down to a number in total

[full lark]

Do they menction this in the new paper?

[grizzled palm]

You're confusing FLOP/s with FLOPs.

[full lark]

But not per second, right?

Maybe that´s the catch (or not, english is not my first languague 😁 )

[near silo]

all I know is this number doesn't make any sense, unless they are counting total compute flops/s x seconds run

[grizzled palm]

But literally no one here understands or cares about the paper details. We just want it to run on our PCs

[grizzled palm]

There could be a typo in the paper, or you could be misreading.

[full lark]

Where did they menction it? @near silo is it on the paper or their twitter?

[trim magnet]

we have a few reddit university graduates here that may care about it

[near silo]

finding it really fast, just a sec

I closed it

yeah, ok, they say it on page 12, first paragraph under 6. Conclusion

[full lark]

Aaa thanks, i´ll see it rn

[grizzled palm]

Yeah I'm at work or else I would at least open the paper. He's obviously just confusing FLOP/s (operations per second) with FLOPs (operations, plural). E.g. 60 teraflops for 1000 seconds (less than an hour) is 60000 teraflops.

[near silo]

they claim it has used 5x 10^22 flops

which is 50,000,000,000,000,000,000,000 FLOPs

[grizzled palm]

That had to be a typo then?

[near silo]

I myself already said that's likely what they were meaning

"Finally, we performed a scaling study of this combination up to a model size of 8B parameters and 5 × 10^22 training FLOPs"

[grizzled palm]

tera is 10^12, so that's 10^10 seconds

[near silo]

10^10 seconds is 317 years...

[grizzled palm]

11574 days on 1 H100.

Assuming 100 H100s, that's 115 days.

So they probably had around 150 H100s and did a few months.

[near silo]

Alright, in that case it WAS that measurement of flops/s x seconds

[near silo]

Wait a sec, you missed a 0
10,000,000,000 = 115740.740740740745
114,740 Days (317 years)

so likely 1k H100's

assuming ~115 days

my guess would be around 1.6k H100's for about 2.5 months worth of compute in total

@grizzled palmoh wait, we did the math at 1 Tera Flop, its actually a lot less

[fervent thunder]

any word what type of hardwares will be required to run this or no idea yet?

i mean sd3

[near silo]

there are various different models, but all of them will run on 24GB VRAM or less

the biggest one will run on a 24GB VRAM GPU

[fervent thunder]

wow

time to earn more coin to get that 4090 haha

[near silo]

you could also get a second hand 3090 for less than half the cost and end up with about 80% of the AI perf

even more for LLMs

for 1 H100, it would take 2,083,333,333 seconds to reach 5 x 10^22 FLOPs, which is 57,870 hours
Which means 2411.25 days nonstop on an H100, or

24 days on 100 H100's... Which seems shockingly low @grizzled palm

[fervent thunder]

well now we just need the release. hopefully in the next month but maybe that too optimistic idk

[near silo]

thats likely not actually that much money, just a sec

lets say based off run pod (I know this is not a very good benchmark), you can get 8x H100's for 37.52/hr

Thats 12.5 of these systems, so 37.52 x 12.5 = $469/h

469 x 24 hours x 24 days = $270,144 around

Thats very little compute cost, honestly. Very very interesting. Now I am sure they actually get it for a lot less due to guaranteed renting time and bulk ordering

[full lark]

I hope nvidia will add more vram to its 5000 series, 16gb is now not enough for mid range

[near silo]

yup

[fervent thunder]

when is 5k series due?

[full lark]

They didn´t announced it yet, i think

[near silo]

later this year

[fervent thunder]

oh ok

[near silo]

if they follow their normal release, at least

its usually every 2 years around October-December

1 things for sure, no consumer grade GPU will be able to train the higher end SD3 models. Likely not even a chance of making LoRA's for them

[fervent thunder]

yeah not if they need 24 gig! wow. i think 4090 only thing even has this

or is 4080 too idk

[near silo]

3090, 3090ti and 4090 are the only NVIDIA consumer grade GPU's with 24GB VRAM

all of their other GPU's are 16GB VRAM or less

[fervent thunder]

oh nice. i thought only 4090 was that much

[full lark]

😭

[near silo]

AMD has many more high VRAM GPU's, but they are far less VRAM efficient and slower for image gen, so they aren't an option either

if they can get ROCm to work better, man, the 7900XT would absolutely dominate the budget AI scene

[full lark]

I hope that i can run it with my 12gb 3060, or i´ll have to buy a second hand 3090

[near silo]

$800 for a 4080 performing GPU with 20GB VRAM that uses like 150 watts less power

[trim magnet]

dont worry im sure they will release new scripts so we can train on cosumer gpus

[fervent thunder]

i have 3080 12 gig so i can run a lower model maybe

[near silo]

They are releasing several different sizes of SD3 models, from 800M params (1.5 size) to 8B params (3x bigger than SDXL)

[near silo]

For the smaller models, for sure. But there is just no way a 24GB VRAM GPU will be able to train an 8B param model/LoRA for one. Its just not possible unless they quantize down and lower the quality considerably

24GB VRAM isn't even enough to properly finetune SDXL, which is a 2.6B param model with 2 text encoders. SD3 8B is an 8B param model with three text encoders, one of which is several billion extra praameters on its own

T5 XXL is almost double the size of all of SDXL on its own

[sudden ruin]

Oh I skipped through it but must have missed that part

[fervent thunder]

they give any hints for a release?

[pine fiber]

34 seconds on a 4090 💀

[sudden ruin]

Its done when its done is all we know

[fervent thunder]

lots of cool stuffs going on with ai

[sudden ruin]

I mean has there ever been a silent week in AI since the release of SD

[fervent thunder]

haha true

[full lark]

It looks like people would have to use at least a6000 to tweak it (unless nvidia adds more vram to the new GPUs)

[sudden ruin]

We can pray

If they dont I have no reason to upgrade at least

Would be kinda pricey, one 4090 for work and one for games

[trim magnet]

yea just wait maybe they will release scripts that can do it on 24gb

[near silo]

bro think's they are gonna release scripts to let a 24GB VRAM GPU train an 8B param model

[trim magnet]

its ok maybe one of the engineers will release something good

[near silo]

its just the price of these models increasing way faster than consumer hardware

[sudden ruin]

Someone surely will develop something for lower end hardware, also since the model isnt finished yet, maybe they will have some breakthrough with VRAM reqs. A man can dream

[trim magnet]

gonna have to ask one of them tho i dont see any here

[near silo]

they didn't for SDXL, and its like 3x bigger than SDXL. I am just saying be reasonable. Don't get your hopes up for something thats almost guaranteed to not happen. Instead be pleasantly surprised on the 0.001% chance that something DOES happen. thats my approach at least

[honest spear]

I think it'll run normally on 8GB, but without T5 XXL model, that thing adds 19GB to the VRAM requirement, just for text😮‍💨

[full lark]

Wait the T5 is just for text?

[uneven wigeon]

they have smaller versions down to 800m, and they will still benefit from tne new arch, cog captions, 16 channel vae etc... even the small ones will probably need to run with TE's frozen on 24gb cards tho for full finetune, at least the big fat T5

[near silo]

I do think very small LoRA's for 8B might be possible on 24GB VRAM. Like maybe rank 4 or 8

[trim magnet]

yea i thought the small one was gonna run on 6gb but just gotta wait and see

[near silo]

exactly. I don't think there is hardly and hope for any training with T5 enabled on ANY consumer grade hardware on any of the SD3 models. Its just not efficient enough for consumer hardware. Could be amazing for professionals and businesses tho

[near silo]

honestly, I could see the 800M running on like 4GB VRAM (without T5 of course)

[honest spear]

T5 = text + a little bit better at very complex images. Aesthetics don't change. It's in the paper released

[near silo]

you can already run SD1.5 on less than 4GB VRAM, and Sd1.5 has a few more than 800M params

yeah, T5 seems like an additional little toy for people who have workstation cards/huge funding. Not something any reasonable consumer should even worry about

My biggest hope is that SAI finally started using a better training dataset (which it seems they have) as their horrible dataset is what held back SDXL so hard. I have spent months training 100's of LoRA's for SDXL and found so much out about what it responds better to

[versed yoke]

You candeactivate t5

[near silo]

yeah, I know. I am saying its not for us

[uneven wigeon]

for art styles etc probably dont need to tune the T5 anyway, the smallest openai clip is also the one that handles art styles best probably due to being trained on a really good dataset, so might get away with freezing the other 2 when finetuning.

[versed yoke]

Its for me

[near silo]

yeah, that seems like a reasonable assumption to me

[versed yoke]

Clip alone is justnot very good

[near silo]

yeah

[full lark]

Wasn't this something openAI did with dalle3? and also have human-written captions

[mortal mural]

So im doing a img2img batch for a video but I want the prompt to stay consistent instead of flickering, can anyone help?

[versed yoke]

And ther is a 75 token limit because of clip

[near silo]

I am sure they more likely made an in house auto captioniner similar to a state of the art COG

[near silo]

yeah, after 77 tokens, prompt adherance lowers. T5 has a 500 token limit IIRC

[versed yoke]

I wonder how it is with sd3 it has both

[near silo]

I am personally in the boat of hoping that SD3 comes out looking much better than it is now. I think training on it is gonna be one hell of a task, and the less stuff that needs to be fixed, the better

[pine fiber]

on what

[versed yoke]

But i dontlike astetic fine tuning

[versed yoke]

?

[pine fiber]

sdxl?

[full lark]

In my opinion, it should have an aesthetic score rating on its captions, something like Ponydiffusion or Midjourney

[near silo]

They did that with SDXL, and I don't really think it helped anything. Granted, SDXL used a pretty bad data set, so maybe that was more of why it didn't work

I don't know, something just feels off about SDXL. Like I don't think I've ever seen it ever replicate any of their demo images in terms of quality, not even remotely closed. And looking back at the original demo images for SDXL, I think those images look better than what they're showing for SD3 right now, which is not exactly the most promising to me

[pine path]

you can use multiple gpus

[near silo]

as somebody who uses multiple GPU's for AI, it is not anything to aspire to do

[pine fiber]

but damn sd3 inference looks really slow

[full lark]

Yeah they should show images that show the "maximum potential" of sd3, in civitAI there are some sd1.5 images that look better

[pine fiber]

for my system it might be like 10 minutes per gen LOL

[near silo]

and I have not found a single trainer that can actually utilize more than 1 for SD

[bleak matrix]

Good morning, everyone!

[pine path]

yes, for SD

[bleak matrix]

How are we all today?

[near silo]

good morning

just discussing thoughts on SD3

[pine path]

guess what's easy to do for transformer based architectures?

[pine fiber]

I think they will be adapted

yay scaling!

[pine path]

you'll very likely be able to do tensor parallel sharding on SD3

[near silo]

that could be promising at least, though I am sure that will slow it down even further, and its not exactly fast as it is

Also, Joe Penna doesn't work at SAI anymore, so I likely won't be getting the early access to SD3 that I was supposed to

oh well, seems like a fair trade for him not being around anymore haha

I think the community was pretty burned with SDXL, so I think the SD3 reception this time around is a little more luke warm

or should I say more cautiously optimistic

I just hope that SD3 can do better photo-realism out of the box than SDXL, cause I have spent months fixing SDXl's realism and the dozens of issues trained into it. Having a cleaner slate to work with from SD3 would be worth it on its own to me IMO

I think the fact that there is still a pretty big lack of any really good realism models for SDXL is a bit of an architectural limit, or maybe an approach limit, but it definitely isn't as receptive to fixing its baked in issues as say, SD1.5

[full lark]

Am I the only one who thinks that they should train a model ONLY using real photography? or it would be unefficient to do? (also more costy)

[near silo]

(not to say I prefer 1.5, as I very much do not, just an observation)

[near silo]

They easily could, and it would likely be fantastic, but it would also be too locked down to a demographic that most people aren't interested in. They release general models that can be turned into decent at any specific thing rather than focusing on being extremely good at 1 thing

[full lark]

Are you sure the problem isn't that almost all models are a merge of other models?

[pine fiber]

ill wait for nsfw finetune... for the fuzzy times...

[near silo]

for SDXL? Thats not really the case. I do a lot of model training and work with a lot of model trainers. I have worked a little behind the scenes with Juggernaut, as I am partnered with RunDiffusion, and I can ensure that they train their stuff, not merge it

[full lark]

Almost all "photorealistic" models look...idk too plastic, like some anime images were in the training data or something, or so many asians XD (withouth offense to asians)

[near silo]

I am currently sitting on several different versions of a realis mmodel that absolutely wipes the floor with Dalle3, MJV6, all fo them. But I have spent months trying to fix some fundemental issues that come with making SDXL good at realism, which have stopped me from being able to sell this model/open it to the public

[near silo]

My model specifically targets fixing that issue, and its done by far the best job out of all of the models I have ever used, but it has some pretty severe issues that I haven't been able to fully solve (been making much better progress since I restarted recently)

popping over to #🏞｜general-with-images for some example images, if you would like to see

[full lark]

Sure 👀

[buoyant moss]

Reading the paper now

Very interesting, although some choices are confusing

5x the cost of SDXL is going hurt thou

34 seconds with 50 steps with 8B SD3 on 4090; SDXL on the same setup for me takes around 7 seconds

[grizzled palm]

You're using tons of optimizations to run SDXL at those speeds. XFormers and fp16 are the biggest, but you also have faster sampler than DDIM. I'm hoping it'll be ~2x slower than SDXL at worst.

[ocean bone]

What about an RTX 6000 ADA? Do you think it would be enough to train decent LoRAs?

[near silo]

48GB VRAM should be enough for LoRA's, yeah

IDk about a full 8B param finetune

[rose sedge]

Is Stable Diffusion open source or only free to download? Cause in youtube someone told me it isnt open source though everyone said it was.

[ocean bone]

probably not. But I'd be content to do LoRAs for the 8b model

[near silo]

SDXl is 2.6B params and you can abrely do BS1 full unet tuning on 24GB VRAM, and SD3 max is 8B params, so assuming the same scale, it would likely need more than 48GB VRAM

[rose sedge]

I am not lawer not have time or undertanding of the licenses

[near silo]

different models are under different licenses

[rose sedge]

What about sdxl and 1.5

[near silo]

SDXL is free to use, Cascade is non-commercial, SD 1.5 is free to use

[rose sedge]

So none is open source?

[full lark]

Open source =/ free

[near silo]

I believe 1.5 is open source, SDXL, I don't believe so

or well, you can edit them both into whatever you want and use them as you wish, if thats what you mean

you can clean out their weights and re-train and use them however you want

[rose sedge]

Hmm.

[near silo]

@rose sedgeIf you can give me some better context as to what you have in mind, I might be able to answer better

[rose sedge]

Well I wanted to know just that

If its open source or not

Cause Emad allways talks about open source

And now it seems sd is not

[full lark]

Technically all of them are open source @rose sedge

[rose sedge]

Though eveyone presents it as it was

[near silo]

AFAIK, all of them are bus SDXL 0.9 and Cascade

[pine path]

SD3 is pretty unlikely to have an open license

[full lark]

SD cascade also is open source but has a non-commercial license

[charred mesa]

SD3 will be non-commercial too right?

it doesn't bother me but that's how its gonna be right?

[pine path]

pretty likely, yes

[near silo]

I'm not sure why it would be non-commercial

if so, that will likely be a pretty huge reason fot it to fail

[pine path]

you'll be able to pay to use it commercially

[charred mesa]

yeah

[full lark]

Anyone know if dalle and midjourney have comertial licenses?

[charred mesa]

but guys

it should be non-commercial so that they can charge services that use SD3

but we can still generate "paid" images, its only non-commercial to host it

not like I think it's ethical to make paid images with Ai image generators but yeah

[near silo]

I see what you are saying, but that would be license to use

not non-commercial

[charred mesa]

idk this is what I remember

[rose sedge]

Who would notice though

[near silo]

notice what?

[charred mesa]

non-commercial to host the model, but the generated images are commercial

[rose sedge]

You use it commercially

[charred mesa]

"Mark.... This is good news."

[near silo]

I work with a company, and we have seen the quote for how much SAI charges for commercial licenses, and let me tell you, it is not cheap by any means 😅

[charred mesa]

I thought it's like $20 monthly

[near silo]

for a single person, yeah

[charred mesa]

if you are using it commercially then you are probably working a lot anyway

[near silo]

not for an entire company

[charred mesa]

yeah for a company lol

some enterprice license or whatever probably

but its good for stability

they deserve it

[near silo]

you need to go up a few orders of magnitude to get to small company level

[charred mesa]

companies can host a model as good as ideogram and DALLE3 and still probably save money in comparison 🤷‍♂️

idk about DALLE3 API price for commercial use

and Stability gets some income

[near silo]

All I can say is I have had multiple companies come to us asking how much it would be to make their own in house model, as they don't want to/cannot afford SAI's prices

[charred mesa]

I personally think they should be getting the money, they trained those huge models using expensive server GPUs and suddenly companies could use it for free? Wouldn't sound good for stability.

[charred mesa]

damn

so its THAT expensive then

🤔

[near silo]

I don't think it should be free, but their prices are insane

[charred mesa]

yeah yeah

I thought it wouldn't be that expensive damn lol

[near silo]

One that I can mention was $350k per year for a single company with a couple dozen employees

[charred mesa]

$350K?!?!?! 🤯

DAMN

[full lark]

DAMN!