💬｜general-chat

[pine fiber]

I noticed those too which is why I think it’s worse

I got so many random limbs and extra finger etc

does horribly with humans

[alpine hull]

Hopefully SD3 doesn't have those issues AND has equivalent or better text capabilities

[shell tendon]

If sd3 is anything like it appears to be, it's much better than ideogram based on my limited testing

[alpine hull]

I wouldn't say much, just noticably

[shell tendon]

If cascade didn't have the leftover latent noise issue I'd call it better as well

[alpine hull]

Though

Like

Ideogram still produces some damn good outputs if you're lucky

Ig that's why I think it could be better?

After all, SD3's demos could be cherry picked

Even Dall-e 2 had cherry picked images

Omg it was NOT anywhere near how it actually performed

[pine fiber]

Emad confirmed they were zero shot

[alpine hull]

WHAT

[pine fiber]

On one of his tweets for the text clothes I think

I don’t doubt it tbh but he does have a good understanding of how the prompts work

[shell tendon]

Sd3 will probably be awesome tbh

I get the anxiety, but think of the jump from sd15 to sdxl.., they've had 8 months to work on this one, which is forever in AI years

[alpine hull]

Still wanna know tho

[cloud plume]

Not ver flashy but I want to test whether it can do "a scatter plot with three points. The first point is at the coordinates (0,0), the second point is at (1,0), and the third point is at (1,1)."

All image models i gave this task thus far have failed 😅

[pine fiber]

especially since it’s a new architecture

[alpine hull]

Ideogram 1.0 isn't finished either, they said they'll keep upgrading it in the coming weeks

[pine fiber]

try with pictogram its free

[crimson belfry]

SD3 is compared to Dall-E 3 or better

[pine fiber]

oh ideogram

[alpine hull]

Ig it all depends on how much both of them improve

[shell tendon]

The diversity in model sizes and what we've heard with cascade tells me they now fully understand the importance of prompt adherence and also training being accessible to more than just people with 24+gb vram

[alpine hull]

1.0 has really good spatial reasoning

[crimson belfry]

We are still in the AI boom

Imagine the industrial age

This is a change of the ages

[trail lion]

as long as it's appropriately woke

[alpine hull]

As someone who really does not know how small text to image models are usually, 8b max for SD3 kinda scares me

[crimson belfry]

Awake not woke

[fervent thunder]

sd3

[alpine hull]

I wonder when invites are gonna start rolling out

[crimson belfry]

Should be soon maybe less than 14 days

[shell tendon]

a day from a day frow yesterday

[alpine hull]

I didn't have much luck with Gemini Pro 1.5 invites, all my friends have 1.5 but not me...

[crimson belfry]

You don't want it

[sudden ruin]

Not cherry picked

[alpine hull]

1M token context window

[crimson belfry]

Google is in the middle of a downfall in the AI sector

[cloud plume]

#🏞｜general-with-images message

[crimson belfry]

Google = AI slavery

[fervent thunder]

what does this mean

[crimson belfry]

Same with Open AI

It means that you are helping them train AI that will be used to turn you into a slave

[alpine hull]

Mfw the model through the API doesn't have the same issues that the product has, and everyone wants to flame Google for their model when it doesn't even have the issues itself

[crimson belfry]

Prison Planet AI is real

[alpine hull]

Someone in the Gemini Apps development team messed up so bad

[trail lion]

meaning? they are enslaving their ai?

[crimson belfry]

Or the creators of it

[cloud plume]

Dalle 3 also failed and also was worse tbo. (So did SDxl). I guess this is a very hard task for ai

[quartz hound]

Hello everyone, do you know those videos where you see plants growing in hyper time lapse? That should work through stable diffusion/Deforum, right?

[crimson belfry]

When you make an AI not recognize white people or say that pedophiles are cool we have an issue

They are clearly coding this AI to go against us

[trail lion]

no idea what that means, but they have layed off over 12k workers last year, and some hundreds this year citing ai

[crimson belfry]

It's no different than government propaganda in the news

We can not allow them to control information.

This technology will only get bigger and they use OUR info to train it, we pay to use a bot that trains on our info

[shell tendon]

It just doesn't understand the prompt, I bet there's one that could generate that with a prompt it can understand

[trail lion]

cant argue with that. the discussion above about closed source being better, I'd argue that we need to keep these open source just to ensure privacy, if no other reason

[shell tendon]

That would take a LLM translating it into a prompt it can understand first I'd think

Unless you have lots of similarly captioned scatter plots in your training set

[crimson belfry]

We need need an online constitution that deals with online entities and AI, Constitutional AI

[trail lion]

but I can see governments starting to lock this down, just for perceived dangers

[crimson belfry]

We need to nip this in the bud asap

Every single update to AI will come with more and more censorship if we do not get a grip of this

[trail lion]

no chance, after the swift thing, etc etc, it's going to be open season for the lawmakers

[crimson belfry]

You are lucky to be able to create a political figure right now and that is unconstitutional.

They will not stop us from creating

[pine fiber]

DiTs are llms

[crimson belfry]

They can not even stop a graffiti artist but they think they can tell us that Trump is bad and we can not make an image of him or sleepy joe?

Enough is enough I do not need a new mom and dad from the government

[trim nymph]

its friday, good day to start the discord thing

[crimson belfry]

Sorry I can not generate that, It goes against the rules citizen

[trail lion]

what discord thing?

[wispy oasis]

i think he means sd3 invites

im getting dissappointed i emailed and used the support form 3 days ago for billing support and no response from stability.ai

[cloud plume]

true scatter plots not commonly being captioned that way makes it indeed not ideal.

_Also the exact prompt i wanted to use was changed by chatgpt4 before It gave it to dalle3 and i just continued using that edited prompt. _

[hexed chasm]

Hi, anyone here know much about Stable Diffusion XL Turbo?

[lone hawk]

forget about it, this is outdated garbage

[hexed chasm]

Last time I used Stable Diffusion was a couple of years ago - v1.4
I'm trying to use Stable Diffusion again for a class project, but am having trouble

[lone hawk]

just wait for SD3

[hexed chasm]

this latest XL Turbo version is too hard to control

[hexed chasm]

Hey, when is SD3 coming out??

[lone hawk]

in 2-3 months

[hexed chasm]

I'd heard about it, so I registered myself on the waiting list

[hexed chasm]

Oh crap, that might be too long a wait for me -- my project is due by end of semester -- probably end of March

or early April

[lone hawk]

yep, we all waiting)

[hexed chasm]

I saw the nice cool new features advertised for SD3, like the better text rendering -- I'd love that

meantime, how the hell can I use XL Turbo?
they seem to have gotten rid of Negative Prompts

How do you control the output when Negative Prompts are no longer allowed?

[chilly tangle]

Guys my Kohya loss is around 0.155-0.16 rn, usually for all my previous Lora they are around low 0.100, is 0.155 considered a normal number

[pine fiber]

its not a very good indicator of training

[chilly tangle]

I will see how the epoch comes out

[fervent thunder]

next year after dalle 4 comes out

[hollow ledge]

what prompt, model etc. do you use to make website mockups?

[hexed chasm]

What's Kohya loss? Is this something to do with Transfer Learning? (eg. LORA)
What kind of datasets do you train on for that? Where do you get these datasets from?

[chilly tangle]

Kohya loss is how similar the trained lora would look like your training dataset

higher the loss the less smiliar it would relate to your dataset

[hexed chasm]

I have a class project to do with Stable Diffusion.
I'm supposed to demonstrate some useful Deep Learning skills in my project, relating to Stable Diffusion.
But so far, my idiot teammates are just typing text prompts into SD XL Turbo and feeling gleeful about it 😫

[chilly tangle]

you get the readings from Kohya_SS GUI while training

anyways it's dropping now

[hexed chasm]

ohh, that reminds me of that French phrase -- Fréchet Inception Distance

[chilly tangle]

[hexed chasm]

Can anybody help me come up with something interesting (but doable) that I can do for my Stable Diffusion project for class?

[chilly tangle]

What kind of project

[hexed chasm]

I need to demonstrate some skills in Deep Learning
But my topic has to be on Stable Diffusion (because that's the topic I was assigned)

I wanted to do some Transfer Learning (eg. LORA) on Stable Diffusion, but I don't know where to get a dataset to use for that

[chilly tangle]

well I do know how to finetune SDXL checkpoints and Lora trainings, beyond that it's above my reach

[hexed chasm]

Has anybody heard of Stable Video Diffusion? It's supposed to be some kind of Text-to-Video application, also made by Stability AI (makers of Stable Diffusion)

[chilly tangle]

#▶｜stable-video-diffusion

[hexed chasm]

That's cool knowledge

[hexed chasm]

so you've heard of it?

[chilly tangle]

It's nothing you just gather the training dataset and feed it to Kohya, that's all

[hexed chasm]

and then what do you produce as your end-product? You make newer/better images with it?

is that the whole point?

[chilly tangle]

Yeah for a very quick example, like pokemon, you can create your own pokemon if you train a pokemon lora

[hexed chasm]

ahh - I saw a tutorial on youtube for that (pokemon)

[chilly tangle]

Yep, it's quite simple

[sudden ruin]

It's not really Text to Video tho, more of a image to video

[hexed chasm]

I saw that LORA-trained model posted on HuggingFace, and tried to use it
The pokemon results were awful

[chilly tangle]

Depending on the dataset

[hexed chasm]

Yeah, I looked for a Google Colab on Stable Video Diffusion, and I found one -- it used an image as the input

It was a painted image of a mountain (Mount Fuji?) and it had fireworks

[sudden ruin]

Yeah it doesn't accept text input so far, but there has been word about a newer version of svd

[hexed chasm]

And running the Stable Video Diffusion code then made an animated image with the fireworks animated

[hexed chasm]

Hey, I'll take image input -- just as long as it can produce a good animation / motion video that doesn't look crappy

I want to be able to demo something decent for my class project

[chilly tangle]

I'd recommend Runway or Pika at current state

[hexed chasm]

my results using text prompts on SD XL Turbo have been lousy
they want me to demo some useful code relating to Stable Diffusion, and here I can't even generate a decent image with text prompts

[chilly tangle]

and maybe sora later on

why turbo though?

Why not normal SDXL checkpoints'

[hexed chasm]

Sora isn't open source, though
At least with Stable Video Diffusion, you can download the model and run it

I'd heard about Google Lumiere first, but that's not open source either

[sudden ruin]

I mean that is more an issue with writing a proper prompt maybe

The model itself ain't bad

[chilly tangle]

True

[sudden ruin]

Maybe don't use turbo but the full base one

[chilly tangle]

Also use a finetuned checkpoint instead of official checkpoint

finetuned models tend to give better results

[hexed chasm]

in desperation, I went to ChatGPT and Google Gemini, and asked them for Stable Diffusion prompts
alright, their prompts produced better image results than I could on my own
but I wish I didn't have to depend on those 3rd party chatbots

[chilly tangle]

Why not create your own prompt

[hexed chasm]

where do I get one of these fine-tuned checkpoints from? HuggingFace?

[chilly tangle]

instead of asking chatbots

[chilly tangle]

Civitai

[sudden ruin]

Check out civitai and browse some images and look at the prompts to get a better understanding

[hexed chasm]

I started out doing my own prompts, and my results were very inconsistent
Hard to control the image results -- lots of weird defects and distortions

[chilly tangle]

well if you have a detailed example maybe I can help you more

[hexed chasm]

do these fine-tuned SD models also have fine-tuned prompting?

like, do they offer better image control capabilities in the prompting?

[chilly tangle]

No just describe follow NLP

[chilly tangle]

they do

[sudden ruin]

Some models feel like they understand your prompt better yeah

Today I discovered the helloworld model which I really enjoyed playing around with

[hexed chasm]

is this Civitai the latest and greatest?
I'd like to use whatever's best, rather than older stuff

[sudden ruin]

Pretty much the only good source right now unfortunately

[hexed chasm]

Also, speed matters for me, since I have to be able to do live demos of the image generation
I don't want anything that takes 5 minutes to render under GPU -- the faster the better

[chilly tangle]

Don't use V5

V3 is better

[chilly tangle]

What gpu are you using?

[hexed chasm]

what's HelloWorld, and why's it good/enjoyable?

[chilly tangle]

For me a RTX4090 gives SDXL output with 1024x1024 pixel under 5 sec

[hexed chasm]

Tesla T4

[chilly tangle]

T...Tesla?

[hexed chasm]

yeah, I'm trying to get something better

[chilly tangle]

Like the Car Tesla?

[hexed chasm]

no, older Nvidia

[chilly tangle]

Oh

[hexed chasm]

It can do CUDA

[chilly tangle]

You had me worried there

Never thought Elon Musk make SD works on Tesla

What's the VRAM for Tesla T4?

[hexed chasm]

So I had a thought about trying to make my own cartoons/anime using Stable Video Diffusion
But I don't know how reliable it is

[sudden ruin]

16gb it seems

[chilly tangle]

16GB VRAM is enough for SDXL

a 1024x1024 should just take few seconds

[hexed chasm]

16GB

[chilly tangle]

Then you are fine

[hexed chasm]

Somebody said Stable Diffusion 3 is coming out soon, with superior text rendering

I registered to put myself on the waiting list

but when exactly is it coming out?

[chilly tangle]

No idea

[honest spear]

I would wait for the moment stable video catches up with sora, but if you think you can get something good give it a try

[chilly tangle]

few months maybe

[hexed chasm]

is Stability AI going to release their own direct competitor to SORA?

[pine fiber]

theyve been awfully quiet about the whole thing

[pine fiber]

not competitor but yes they are working on stable video 2

[chilly tangle]

I do recommend looking into closed source servieces like openai or Midjourney at this stage

[hexed chasm]

or is Stable Video Diffusion all we have to work with for the foreseeable future?

[chilly tangle]

MJ is coming up with text-to-film with next version release

[hexed chasm]

Well, I have to demonstrate something where I can do some actual coding, beyond just typing in text prompts

[chilly tangle]

Why not just teach python then

Coding is not a must using SD

[honest spear]

From what I understood, sora put a milestone that is: video generation is similar to world simulation. Emad said something like stable video will have some of those aspects, but I haven't heard too much

[hexed chasm]

when is this text-to-film thing coming out from MJ?

[chilly tangle]

This summer with V7

[sudden ruin]

When it's ready it seems

[hexed chasm]

well, that's the whole thing - I'm supposed to use Python, because it's a Python course

well, it's a Deep Learning course, where we have to code in Python

[chilly tangle]

MJ is pushing text-to-film release earlier due to the release of Sora, MJ is tend to compete with Sora from what I've learnt

[hexed chasm]

I wish Stability AI would hurry up and come out with a competitor to Sora

[chilly tangle]

I'm worried if they are being funded enough

[hexed chasm]

Do you guys have any experience with Stable Video Diffusion?
Text-to-video is starting to look a lot more impressive than text-to-image (old stuff)

[honest spear]

to compete with Open AI? Stability have 100x less than them xD So I'm happy with what they're giving us

[hexed chasm]

(okay -- image to video)

[sudden ruin]

It's alright but has a long way to go, also limited to rather short clips

Maybe you want to check out animatediff

[chilly tangle]

That's what I meant

You shouldn't ask too much when someone is giving you something good for free

[hexed chasm]

Any place where I can check out samples of Stable Video Diffusion output?
I just want to see what kind of stuff is possible

[chilly tangle]

#▶｜stable-video-diffusion

[pine fiber]

its not that good

[hexed chasm]

wait -- is that a channel / room inside this discord?

[chilly tangle]

Yes

[hexed chasm]

oh hell, never even noticed it before

[chilly tangle]

You just click the link and it will guide you

[hexed chasm]

ok, brb

[honest spear]

@honest spear True, but there are ways for us to help. For example it's possible to donate on github to people making extensions

[acoustic solstice]

hey guys, i am following the example script instructions in this link: https://huggingface.co/stabilityai/stable-cascade#code-example and am getting an error despite operating on linux venv with all libraries and whatnot installed: https://imgur.com/a/eLub0US

[sudden ruin]

That's one for #🤝｜tech-support

[chilly tangle]

Sorry I wish I could help, but do try #🤝｜tech-support

[acoustic solstice]

will do thanks

[winter pike]

specially knowing SAI is one of the few interested in open source AI, i can give constructive criticism but i would never feel entitled enough to demand things

[hexed chasm]

Yeah, you guys are right - the Stable Video Diffusion results aren't very good

maybe it would work better on cartoon images

[chilly tangle]

Runway do provide free generation times for new users, maybe you can check that out

[hexed chasm]

Runway is closed source, though

I can't download their model and use it myself

[bleak maple]

Yo

Where do i generate images

[jovial wraith]

the discord bot is currently down, so your best bet is either run it on your own computer if it's powerful enough, or get a paid service

[bleak maple]

Oh

[jovial wraith]

or use a free service, but those usually aren't that good

[bleak maple]

I gotta run it locally then

Which i have

[bleak maple]

When is it coming back

[jovial wraith]

probably when SD3 releases. not sure tho

[bleak maple]

Ooh

Then it will run sd3 right

[jovial wraith]

i think so, yeah

[bleak maple]

Midjourney going down MUAHAHAHAHAH

[prisma portal]

Discord really be scewing me over with the game im playing

[chilly tangle]

Not likely, they are pushing V7 for an earlier release

[grizzled palm]

Midjourney secretly switches to SD3 behind the scenes.

[pine fiber]

whats v7

midjourney can afford to make a bigger model

[chilly tangle]

New Version of midjourney

Current is V6

[hollow ledge]

which software do you recommend for generating images now? I have used Ivnoke and automatic1111 before

like long time ago

[chilly tangle]

ComfyUI

[pine fiber]

idle

[grizzled palm]

A1111

[hollow ledge]

?

[pine fiber]

python idle

[hollow ledge]

do you have any model that does a great job at designing websites?

[pale latch]

Midjourney wont die. It's from the same guy who did leap motion gimicky hand tracking that was a half solution and not very well implemented. He'll farm venture capitalist money until the SAAS fad runs out and then midjourney will just sorta fade away. No monumental crash. No burning. It'll just fade. Same business strategy

[lone hawk]

isn't it using like a custom trained sdxl model?

[pale latch]

nobody knows. they keep all their tech proprietary and hidden behind interfaces

[lone hawk]

then it confirms this is the case 😄

[pale latch]

i dont thiunk that's really how confirming works. Its okay to say "i don't know". You couldn't know.

[lone hawk]

you know what gives them up? Same prompt understanding as sdxl

[pale latch]

confirmation biases right? you'll see what you look for. sdxl uses a popular open source clip model provided by open-ai.

[winter pike]

i remember sd 1.2/1.3 getting leaked and a few days later midjourney got from generating very LQ images to good ones, not implying anything but yeah, its def a coincidence

[lunar nebula]

Hello, I have a question. Are there any streamers or youtubers which shows full walkthroughs (like lets plays) using Stable Diffusion? I want to see how others work with SD and I think that watching this will help me a lot 😉

[lone hawk]

maybe this is exactly why OpenAI turned into ClosedAI xD

[pale latch]

that was for ai safety. while they are a saas company, i don't think they care too much about other companies in their space. they're kind of doing the blue ocean strategy, where they're building products no body else has. competition isn't their corporate strategy. i think they legitimately care about ai safety and maybe too much. remember when they fired altman and then they all had to quit to get him back?

the old board were a bunch of effective altruists. psychos about ai safety. they want the government to get a nuclear weapons program developed specifically targeting data centers

[trim nymph]

i recall that too.

[pale latch]

emad was one of midjourney's early supporters so he might've even given them access to his training cluster for their own sd fine tunes. point is, we just DONT know. we can only speculate.

[lone hawk]

oh yeah, i remember that situation... damn, anyway i always thought they are amazing at what they do! MVP company.

[pale latch]

nothing confirms anything

[pale latch]

the whole company basically signed on to leave openai if altman didn't come back. haha. its an admirable company culture

[lone hawk]

all for one, one for all

then how did Stability come up with this new SD3 tech?

[pale latch]

we dont have the technical report yet but they mentioned a couple papers they used.

https://arxiv.org/abs/2212.09748 transformers
https://arxiv.org/abs/2210.02747 flow matching

[lone hawk]

yea, transformers, this is it - Sora also using it

[pale latch]

my 4080 has a transformers engine. i hope theres huge optimizations available

[hexed chasm]

what's flow matching?

[charred mesa]

it would be good to get that paper to see the text encoder and stuff

[pale latch]

its in the paper. i can barely understand it, let alone condense it

[charred mesa]

SD3 technical paper would be way more simpler

[hexed chasm]

why not ask Google Gemini to explain the paper?

[lone hawk]

not sure about that, haha

[pale latch]

i'm in canada and google hates canada now so it won't deploy ai services here

we sued them for millions because they were monopolizing canadian news and they got all pissy and started talking about pulling all services from canada

google news was actually turned off completely in canada for a couple weeks

[lone hawk]

more like woke news 😄

[pale latch]

ugh. people using woke in that context. says a lot. says a fucking lot

[lone hawk]

how did they turned in such racist company idk

[pale latch]

yeah. they're the racist one. ... sure. dropping this conversation. can't stand "antiwoke" people. i know what they really are all about.

[pearl ocean]

Imagine working at X

[hallow sable]

canada is a woke nation

[lone hawk]

new SD3 picture at: https://twitter.com/_shadowanderer

#SD3 #🧣｜comfy-ui

so Comfy already supports SD3

[hallow sable]

isnt just a model???

why would it not

[lone hawk]

new architecture and all...

[charred mesa]

SD3 comfyui hmmm

This is great news

so we will have support in comfyui on DAY #1 just like SDXL

they probably use a secret branch that only Stability has access to lmao

[hallow sable]

I try to use SD for game environment and it does a terrible job with buildings lmao

[pale latch]

another dude making fun of the concept of "woke" holy shit. american cult of trump is dumb af. stop with the dumbass politics

[charred mesa]

yeah buildings are mid

[hallow sable]

it is woke

its goes hard aganist the everyday person

[charred mesa]

okay lets not get off topic with politics 💀

[stray knoll]

Hey community, I'm trying to share a custom node to the subreddit, but Reddit is deleting my topic due to some keyword.
May someone review it ? Thanks !

[hallow sable]

reddit sucks lmao

your better off not using that dog shit website that cant even get video players to work right

[lapis urchin]

ive noticed too many details in the positive and negative prompt leads to the images being saturated

is that due to there not being enough sampling steps to go through it all?

[winter pike]

i try a lower cfg when that happens, specially if im using Loras

[hallow sable]

doesnt lower cfg make washed out images

[rich kestrel]

no

[hexed chasm]

Which version of SD are you using?

[lapis urchin]

latest

[hexed chasm]

I've been away from SD for awhile - just getting back into it - testing out XL Turbo

And it doesn't have any negative prompts

I used to use negative prompts a lot before - but now that I've started trying XL Turbo, it doesn't have them

[lapis urchin]

web version?

[hexed chasm]

I'm using it off Google Colab

[lapis urchin]

ah

im locally generating

[hexed chasm]

just running the python code directly

[hexed chasm]

but do your negative prompts work?

[lapis urchin]

im using the dreamshapeer model

and yes

[lunar nebula]

Hello, I have a question. Are there any streamers or youtubers which shows full walkthroughs (like lets plays) using Stable Diffusion? I want to see how others work with SD and I think that watching this will help me a lot 😉

repost since I crashed in a heated discussion. Hope it is OK

[hexed chasm]

I added negative prompts into the code, but they had no effect

documentation said negative prompts aren't supported anymore

[hexed chasm]

Are you able to consistently avoid deformities and distortions when you generate images?

[lapis urchin]

yes

[hexed chasm]

I find that half or more than half of my images have some kind of deformity or distortion

[trail lion]

what's the point of turbo

[fervent thunder]

How do you generate widescreen images without getting chernobyl looking characters?

Every time I generate something that's not a 512x512 or something similar, and go with a wider aspect ratio instead, i get very weird looking characters with horrible anatomy

[trail lion]

quality > quantity

[rich kestrel]

many different methods

outpainting is easiest

[fervent thunder]

What is that

[hexed chasm]

turbo is fast

turbo is fast - it has fewer diffusion steps

[trail lion]

yah, ppl need to get over this obsession with speed and it/s

[rich kestrel]

this

[hexed chasm]

I always get distorted characters regardless of what size I use

[rich kestrel]

double this

ppl put too much emphasize on speed, like, why do you need to gen 10,000 images

rather generate 5 of good quality

[hexed chasm]

for the purpose my stupid class project, which involves doing live demos, I'm glad turbo generates fast

but it makes so many distortions and deformities - feels like garbage

[meager sage]

hai

[wispy oasis]

hi

[charred mesa]

hello

[wispy oasis]

i just purchased a new pc today to run stable. the parts are on the way. i bought it for the only reason to generate 10,000 images not 5.

i need 10000 images for a project i am making

[meager sage]

[charred mesa]

yes poopmaster

[wispy oasis]

would you rather have 5 dollars or 10000 dimes

[wispy oasis]

i think this will only drive petter performance. i only bought this gpu today because stability api is too slow and can not generate the images i need reasonably fast

for a time constrained project

[lunar nebula]

Guys, I have another question. I didn't came with a solution or better I ask how and why. A friend wanted a picture of a girl smurf with a bib. She is a kindergardener ... So I prompted "a smurf girl child with a bib" using dreamshaper. But I only get pictures of real girls in blue dresses. Since I try to understand and prompt correctly I wonder where I have to work to get a real smurf out of the AI ...

[rich kestrel]

use a lora

and an animated model

[hallow sable]

2560 x 1440 takes me about 2 min to generate

[pearl ocean]

4K when

[acoustic solstice]

hey guys, i am not having an issue generating images at 1024/1024 but when I go to try to create higher resolutions the code still runs but no image is generated, why is this?

[lunar nebula]

Thank you. What model would you use? Dreamshaper was my goto so far. But when I browse civitai I see a lot for Anime, but I couldn't tell why I should prefer one of this over another. I try to understand the backgrounds (thats why I ask for real examples since it helps me seeing the basics).

[hallow sable]

everyone keeps training models on anime and porn...

[split tapir]

question if I have a image that was taken with low quality camera and I make it higher quality some how?

like I guess the pixel count or w/e is low idk much about photos

[pine fiber]

what

thats called upscaling

[split tapir]

yeah I tried upscaling but the quality remains the same

[winter pike]

maybe something like this but its not very convenient atm: https://github.com/Fanghua-Yu/SUPIR

[split tapir]

damn thats exactly what I need but the requirements are insane

I have low quality input image and want it converted to high quality input image 😦

[lunar nebula]

Its felt like 99% of all of it, yes. But what model should I use and what should I do to get what I want? A colleague of her used Microsoft Copilot and got some results. I want to understand the process but it really feels that I have almost no controll what I get when I want something specific ...

[winter pike]

yes it is, im waiting for their online demo for that same reason

but it gives cool results, better than gigapixel

[pale latch]

the fast models aren't even that impressive. less visual quality than their default counterparts. there are huge tradeoffs. more finnicky with settings too so you're operating in a very narrow settings window.

right now, i 4-5 seconds per image. .5 second per image isn't going to alter my workflow or accelerate it at all. We've peaked for speed for that use.

Real time generation will be cool but new purposes will be used with that speed. It won't benefit compositional artists still. The speed domain is just like people trying to get 300+ fps in a game they're running on a 120hz screen. Its just a flex. Like most spoilers on street cars

[deft nymph]

Hi all !
Anyone here tried SD3 in preview?

Just want to know how accurate it is towards prompt understanding.

[pale latch]

just whats on twitter. no early access setn out yet

been a week since the announcement that it was coming to preview soon. I think they announced the announcment too soon

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

[deft nymph]

Okay 👌 thx for the info
At first sight it seems pretty good based on their insta posts

[grizzled palm]

4K in 26 seconds with Stable Cascade on my 4090.

[rich kestrel]

the virgin diffuser in native 4k

[pale latch]

i only see the logo for it on their insta? you mean x posts? theres a bunch in the #SD3 tag

[rich kestrel]

vs the chad upscale from 512 to 4k

[pale latch]

think you got that backwards there bud. the chad likely has the 4090. that's just how pcmr works. sorry not sorry

[rich kestrel]

the implication is that native 4k is just going to be ass

give it another 6 years

upscale is da wae

[pale latch]

4090. bud

figure it out

[rich kestrel]

gpu power has nothing to do with it

[deft nymph]

In SD instagram

I am just interested as a midjourney user about the capabilities of SD3

Their post seems promising tbh

[pale latch]

https://www.instagram.com/p/C3qWXM_vtXE/ thats the only one i see for sd3 on their gram

[deft nymph]

https://www.instagram.com/stablediffusion?igsh=MTJtamQ2bzM0dTJkbQ==

[grizzled palm]

I'm using a native-rendered 4K desktop background. Landscape of mountains and a lake.. No img2img / controlnet / upscale / tiling. Just prompt -> 4K. Made it the day I got Stable Cascade on my PC. Will upgrade when I get SD3 on my PC.

[pale latch]

thats a you situation

[rich kestrel]

landscapes are easy as balls and probably even easier to uspcale

[deft nymph]

Or am I wrong, these latest insta posts doesn’t come from SD3?

[rich kestrel]

sorry but 4k native is just a dumb idea. my opinion after all

[pale latch]

youre right. that's a different instagram than what got linked on their blog post about sd3

[grizzled palm]

Go to general-with-images and I will smack you in the face with gorgeous Stable Cascade 4K. Rendered in 26 seconds. Your opinion is a dumb idea.

[deft nymph]

I am noob with how custom models works but when SD3 will be in full release, does train models will update to it?

Maybe I misunderstand some words, but like, comfyUI will be updated for SD3?

[pearl ocean]

I got a 4080, how long would that take?

[grizzled palm]

I've never owned a 4080, sorry.

[pale latch]

doing it right now

[pearl ocean]

[pale latch]

shit. comfyui doesn't report render times

[grizzled palm]

LOL I literally had a timer running on my phone when I was testing. 🤣

[pale latch]

shit i refreshed and i don't know where it saves. i'm so lost in comfyui. fk

[pearl ocean]

yesterday I tried to set it up, it kept erroring

[pale latch]

need to have the newest most up to date version. i used the workflow examples for cascade comfy made

[pearl ocean]

😮

I do

just errors when running the promt

[pale latch]

¯_(ツ)_/¯

[pearl ocean]

¯_(ツ)_/¯

[viscid cairn]

hii

hi / hola / oi / hallo / bonjour

Someone who has experience in sound computing, spectrography, sound spectra, etc. I am solving a case with hidden sound codes and cesar and ascii codes.

[pine fiber]

Emad posted a pic with higher resolution

2688x1536

multiple

I wonder if they are upscales or sd3 can really go that high res

[radiant tusk]

Have you guys found a controlnet setting that matched facial expressions the closest?

[lone hawk]

he posted without #SD3 tag, that's not a good 😦

[pine fiber]

he replied to me and said its sd3

but with an upscaler

"sd3 => stableimage"

[lone hawk]

phew, hope it's still coming 😄

[charred mesa]

those images look good for a base model though

[chrome violet]

Heya, just wanting a bit of help about models. New to stable diffusion and currently the next step is putting in a model which is Rev animated for me and just wanting some help to how to put it in the section models since I can’t seem to do so. Shoot me a dm and thanks in advance :))

[fervent thunder]

Is it ok to have 512x512 starting resolution if I want a high resolution image at the end? Can the upscalers do that much work

And what is the best way to upscale for quality

[trim magnet]

hi-res fix latent

[honest spear]

totally ok if it's the resolution at wich the model was trained. It's usually 512-> highres fix to get at ~1024 or what you like -> x4 upscale /tiled upscale. Everyone has his own method

[fervent thunder]

Yep that's why I wanna do 512

It says it's optimal for the model

Is it better to save GPU power and get more images by not using highres fix

And then pick out the best ones and upscale them after the fact in img2img?

Or will highres fix and then upscale with img2img produce better results

[trim magnet]

u just use seeds of the imgs u like then hi-res fix

[fervent thunder]

How do seeds work

[honest spear]

Yes, probably better to have a batch of images generated, then pick the best result and do an img2img. Highres is foundamentally an i2i process.

[fervent thunder]

Yeah it just adds so much time to each image, most of which im deleting

[pale latch]

kind of. it boosts off the latent before finalizing it to image form. i2i has to turn it back to a latent

[fervent thunder]

https://youtu.be/EmA0RwWv-os

Is this a good method?

[trim magnet]

depends,if u want latent upscale u have to do it in txt2img or if u want a very big upscale u can do img2img with controlnet,latent upscale works better on comfy with the NNlatent upscaler that doesnt destroys the img like it happens in a1111

[fervent thunder]

What is the difference between latent and in img2img

and i am in a1111 so i guess that answers that question

[pale latch]

whats the nnlatent node all bout?

https://github.com/Ttl/ComfyUi_NNLatentUpscale found it

[honest spear]

Latent space is the form that the model uses, so until you remain in latent space you should have no loss of data. If you start decoding to image, then back to latent, for example when doing i2i, you'll have to pass through a decoder (the vae) and lose a tiny bit of data

[fervent thunder]

Ahh

That explains what I saw

[pale latch]

i like to pretend i know what latent is, but i'm not sure i do

[shell tendon]

single most annoying thing about pretty much every SD-anything: the forced downloads that stop the program unless you force quit it, that don't specify what is being downloaded or where it's going, when it's probably something you already have a copy of >_>

[pale latch]

been pretending so long its ahrd to tell

[shell tendon]

webui forge, after updating, is now trying to force download realisticvision5.1 for whatever reason

i already have four versions of it but it wants that one

annoying as F

[trim magnet]

and then the dl gets stuck

[shell tendon]

yeah

for gods sake give us an option

i'm not on the fastest connection and this crap is eating up my HD space

every 2GB download = another 10 min waiting

i don't need a default checkpoint donwloaded, idk why the hell forge has added that as a thing

and... the update script uninstalled pytorch 2.1.2, installed 2.0.1, and now its complaining its the wrong version on launch lol

jfc

[pale latch]

are you sure it's not some extension? i use forge and it's never auto downloaded models for me

sounds like a layer8 issue. good luck

[shell tendon]

forgot it wiped my links to other directories, so that's fixed now, but yeah

wish that in general this stuff would give you a link and a destination directory and then give you the option to autodownload or to do so manually so you can do it without lokcing up the program

[buoyant moss]

SD3 wen

Me want

[wicked nebula]

Hi everyone. I'm new here and don't quite know the rules but I need some help. I used Stable Diffusion at the end of last year to create some cool retro comic-like art and used a version of Stable Diffusion that had comic as a style option. These days, I can no longer find that option. Can anyone help me out? Thanks.

[winter pike]

the bot is down right now, we hope its back soon

[wicked nebula]

@winter pike thanks

[thorny osprey]

How much better is SDXL vs SD1.5?

[shell hazel]

most people prefer 1.5 still

[thin oriole]

Quick question. I signed up for the SD3 wait-list, does the Discord ID section mean my username or the User ID that's a random string of numbers?

[pine fiber]

both should work I think

I would put username

[pale latch]

i'm certain they just fucked up the language there and mean username not discord id.

even though, discord id is what the numbers are called

if they actually require the hidden behind developer mode discord unique id, i guess i'm not getting into the preview

[thorny osprey]

Ty for your response. I hear XL and XL Turbo are faster but require more VRAM. Will a checkpoint from XL or XL Turbo work in 1.5?

[shell hazel]

no they don't mix

[pale latch]

1.5 is a model architecture. you don't use xl with it.

popular UI's will load all known models based on stable diffusion's base models

the ui's are software that load the models you can get off huggingface or civit.ai

[radiant tusk]

How do you get a generation to instant close?

When I click interrupt it always has to finish first

[pale latch]

close the terminal

alt f4

[split lance]

Hey hey does anyone know if the v2alpha inpainting endpoint can take width and height parameters?

[open elk]

Hey there. I have a really stupid question. Where do I go to create an image? I haven't been in this group in a couple of months, and it looks like verythign has changed. Wehere will I input my prompt? For some reason I can't find anything about this in the "start here" section either

[visual warren]

I think they have the bot down while they are upgrading it to the next version, but im not totally sure on that one. Been meaning to ask myself honestly.

[open elk]

Thanks so much. I thought I was losing my mind...

[buoyant moss]

Just hoping they launch SD3 (at least a web demo) next week

[hearty axle]

hey whats the command to make a photo realistic image and which channel is best for it , I forgot the command because i used it months ago . is it #dream or #imagine ? \

[nova zodiac]

none of the above - bots are down #1047610792226340935

[radiant tusk]

How small of a face can adetailer detect before it becomes undetectable?

[pearl ocean]

I feel SDXL might be giving me better details than SD

[stone stirrup]

we all do... there is a waiting list, but only experienced creators will get in I am afraid

[pearl ocean]

soon, use SDXL for now

[hexed chasm]

Why do most people still prefer 1.5 ?
What does 1.5 offer that makes it more preferable to XL ?

[hexed chasm]

only thing I like about XL is that it's fast -- but I find the image output to be lousy

[jovial wraith]

it's faster, has more loras, and i'm pretty sure it's easier to train as well. among others

[hexed chasm]

1.5 is faster than XL? why?
XL is said to have fewer diffusion steps, which makes it faster

[pearl ocean]

so SD is better?

[hexed chasm]

"has more loras" - what do you mean there?

[jovial wraith]

do you know what a lora is?

[hexed chasm]

Low Rank Adaptation

a means of fine-tuning models efficiently

I take Deep Learning full time in school

[jovial wraith]

exactly. they have to be trained for a specific model. so if your lora supports sd1.5, it doesn't support sdxl.

and vice versa

[hexed chasm]

but when I do LORA, then I don't use somebody else's web gui interface that they already coded
I just write python and run it myself directly

[pearl ocean]

have you even met Lora

[hexed chasm]

haha

I have to do a class project, and I've been assigned a topic of Stable Diffusion

[wispy oasis]

hi im curious, what level of school

[hexed chasm]

At first, I thought this topic would be fun -- but now I'm hating it

[hexed chasm]

grad studies

[wispy oasis]

cool

[pearl ocean]

Phd when

[hexed chasm]

I feel like an idiot doing this Stable Diffusion topic
I have to make a demo for class, and I don't know what to do
so far, I'm just writing text prompts -- which feels ridiculous for a coding class

[pearl ocean]

write how the nodes work

[hexed chasm]

what do you mean by nodes?

[pearl ocean]

how the language model works

[hexed chasm]

well, I assume it's using Word2Vec or something like that
I already have another class for NLP (natural language processing)
But in this class, I've been given the topic of Stable Diffusion
I thought it would be fun, but now I'm feeling like I drew the short straw

Everybody else is doing Computer Vision stuff
I got assigned Stable Diffusion, and thought it would be cool because it's generative

[jovial wraith]

doesn't SD have some sort of computer vision as well in the form of CLIP? in the img2img tab?

[fervent thunder]

theres clipvision and unclip

[jovial wraith]

ah, i see. sorry i don't know a lot about this kind of stuff

[nova zodiac]

only thing clip is used for in img2img is turning the prompt into tokens

[hexed chasm]

CLIP is actually the text encoder for SD, as far as I know

[jovial wraith]

either i'm misinterpreting the description on github, or it can do both. not sure, again, i don't know a lot about this stuff

[hexed chasm]

I assume 1.5 still allows negative prompts.
When I tried using XL, it doesn't seem to support negative prompts anymore - at least that's what people told me

How can I use negative prompts in XL ?

[jovial wraith]

what ui are you using? and in most cases it's good to leave the negative prompt blank with XL models. at least that's what i'm told

[hexed chasm]

I'm just running the python code directly inside Colab - not using any GUI
I'm supposed to code something for this class project, so I don't want to use any GUI stuff

But my problem is that I can't find anything to code -- I don't know how to modify, improve, or add to SD in some useful way

Maybe I should just do LORA

[jovial wraith]

again, i'm no expert, but using SD without a ui is going to be difficult. maybe you can give ComfyUI a shot? it's really complex and supports custom nodes if you know how to make them. that could be a way to contribute.

[pearl ocean]

my question is, is it possible to make money off SD images?

[wispy oasis]

yea

[hexed chasm]

For your LORA experts -- can you tell me if LORA can help SD make better comicbook art images?

[wispy oasis]

sell images as nfts

[hexed chasm]

Can you all describe to me why you use LORA? what kinds of images do you make? why are they better, in your opinions?

[wispy oasis]

i dont use lora

[hexed chasm]

what is your preferred way of using SD?

[pearl ocean]

Is crypto even still thing?

[hexed chasm]

what kinds of prompts do you use?

[wispy oasis]

well so far i have only used the API on stability api. i have orded a new computer parts yesterday and now they are on the way for stable diffusion. when i download sdxl myself i will be using CLI and scripts only, no web UI.

[pearl ocean]

4090?

[hexed chasm]

and what kinds of prompts are typical prompts from you?

[wispy oasis]

well its supposed to be an animal in the same pose with different traits, thousands of images

[hexed chasm]

and what does your prompt look like?

[wispy oasis]

4070

[wispy oasis]

hybrid protocols are new and combining tokens and nfts in to one

[hexed chasm]

Can you give me a sample prompt?

And what version do you use? XL? XL Turbo?

[wispy oasis]

my prompt format is still being worked out. right now im trying different prompts but it is usually something like "a cartoon portrait of a bear, wearing <type of hat>, and <type of sunglasses> on a <color> background where these <traits> are randomly chosen

[wispy oasis]

sdxl

[hexed chasm]

thanks

[wispy oasis]

but im waiting for my pc to get here before i will use it on my own pc

[wispy oasis]

to generate similar images i use the same seed after i see a seed i like

[pearl ocean]

I love 64gb of RAM.
If I knew I was gonna be using AI on desktop in the future, 8 months ago, I’d probably have gotten a 4090 over a 4080.

Even though it was like $1200 more at the time

[low moon]

Kind of a weird question but I think this may be possible: Is there a way to put promts in some kinda txt file and reference them in the promt box like we refrence loras and embeddings?

so that i dot clutter my promt boxes with 500 words

[low moon]

Oh its the styles thing.

I sure hope comfyui allows this too

[carmine pecan]

to generate similar images i use the same seed after i see a seed i like

[hexed chasm]

How long a prompt can you supply to Stable Diffusion? What version are you using?
I was recently trying XL Turbo, and it says it won't accept more than a 77-token limit

[hexed chasm]

Does that seed only work for multiple runs during the same session?
Or can you come back a week later and use that same seed, and still get the similar results?

[pseudo jetty]

so i just updated to 1.8 on a1111 and i was wondering if anyone new about any ui settings that allow me to revert back to the old model card layout, instead of this new one without just downgrading?

[tribal void]

what is the file size of SDM v2

[hexed chasm]

so you're using 1.8 -- someone on here was telling me before that v1.5 is the best
why do you like 1.8 ?

[pseudo jetty]

i dont really care for the new update for a1111 tbh, i just like to keep my A1111 updated to the newest commit and see if any new interesting features, but ill be reverteing back to 1.7 for the time being.

but to me knowledge a1111 1.5 isnt anything special? i think its the farthest you can go back while still having SDXL support

unless they meant SD 1.5 in that case thats a completely different thing

[sonic cosmos]

hi

iwanted to prompt and i got this

RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument index in method wrapper_CUDA__index_select)

what does that mean

[fervent thunder]

Whch is better, latent or 4x ultrasharp for highres fix?

[hexed chasm]

So is SD XL Turbo the latest and greatest version? Or is SD XL Lightning newer?

[sudden ruin]

try asking in #🤝｜tech-support

[shell tendon]

You could always just build a second PC with a 4090

[shell tendon]

Seeds are forever

Seems that sd3 deadline went whooshing by like a freight train and derailed off the edge lol

[fervent thunder]

rtx 50 should be comin soon

relatively

hold out and see what they got

$1200 for 4080->4090 seems extremely unfavorable

it's pretty much the same in my country

+1000$ for each jump between 4070ti->4080->4090

[lone hawk]

Maybe they're not going to release it at all, i mean why would you give such amazing image generator for free? From business standpoint it makes no sense, so yeah.

[fervent thunder]

They could just go chatgpt route

Keep previous one free, make the new one a monthly membership

But I think it being free is very important

[lone hawk]

i thinks they turn it into Dall-E 3 - Midjourney, so yeah, no SD3 for us i guess.

[fervent thunder]

It doesn't compete with midjourney

imo

It's the free part that draws people in to make their own models off of SD

From what I can see

And THAT is what makes it good

Because from what I saw about SD3, it does not hold up against midj

So if they are both paid options, people will just take mid

[lone hawk]

Free works only on crap models like Cascade, SDXL, etc... not the case with SD3

[fervent thunder]

What makes mid so good anyway, like how is it so good

How has it managed to stay on top

Every time I see images generated from it, I'm just wowed

[lone hawk]

Guys, get this - they showing those SD3 pics on twitter just to get an attention from investors!!!

[hoary sail]

On a white paper, a large Chinese character is written with a brush.

[trim magnet]

stop spreadin bs here m8

[lone hawk]

i bet on this, you'll see 🙂

[trim magnet]

bet what

[lone hawk]

one dollar 😄

[shell tendon]

They'll release it

[shell tendon]

Better get started on the paperwork :p

[winter pike]

its not the first time things get delayed so keep calm

[charred mesa]

^^

they promised it to be open no matter what

[sacred glacier]

a cat

[shell tendon]

Here is the image you requested. #🏞｜general-with-images message

[shell tendon]

yeah, within a year prolly

remains to be seen if it's a real upgrade and what price gouging nvidia pulls on it

tbh if they bump the ram to 32gb i wouldn't be surprised if they also bump the msrp north of $2k

[lone hawk]

oh what do we have here, isn't it just like that closed Midjourney discord bot!? https://pbs.twimg.com/media/GHrCo38WEAA8zUM?format=jpg&name=large

[shell tendon]

that and if you think the 4090 is hard to get now...

[lone hawk]

Open SD3 release is a sweet lies you guys....

wake up!

[shell tendon]

rise n shine, time to make my SSD whine... downloading more fn models >_>

i'm fn addicted

[fervent thunder]

They mentioned it being the biggest leap yet, implying it'll be bigger than 4090 vs 3090

Implying it'll be over 2x faster

If that is true

You will have one hell of a gpu

[fervent thunder]

What is the point of running it on an SSD

[wicked sonnet]

😭 I asked my local ai to introduce themselves and it's up at 1900 words and still going

[fervent thunder]

I got my eyes on the 5090 too as soon as they release it

A100 is as close as you will get I believe

Right now

It has some insane memory

But also an insane price tag

[wicked sonnet]

chip producers must be making a lot of money

[shell tendon]

well if that's that's the case i'll be building a third PC the second it comes out

this is my one and only hobby these days and i don't really spend money on anything else so F it

[shell tendon]

faster load times when switching models

[fervent thunder]

Yeah man get the best u can afford

Oh relaly

I should move my folder to nvme then

Is there any issue iwth moving it from one drive to another?

Do i need to change some values anywhere else?

I got a shit ton of m2 nvme space just sitting there

[shell tendon]

nah, not really

what are you using? comfy? a1111? forge? something else?

[fervent thunder]

a

[shell tendon]

i have two folders for checkpoints, one on my m2 ssd and once my drive starts filling up i shift less used ones onto a HD

[fervent thunder]

give it a cookie it must be really proud of itself

Yeah im gonna do that too

[shell tendon]

read speeds are something like an order of magnitude faster at least

[fervent thunder]

And keep the images on hdd too

Cause I am always switching up models to see how it looks with different ones

If that becomes faster on ssd that's a huge time saver

[shell tendon]

i can load a sdxl checkpoint in prolly just under 2 seconds i think

yep

it's a 1-2 second load time i think

[fervent thunder]

awesome

on my HDD it's like

40-45s

[trail lion]

it's an Entmoot

[wicked sonnet]

😭 I had to stop it at 4k words

[wicked sonnet]

what's that

[thorny sedge]

hi

[fervent thunder]

Can you put it into a txt and send it to me

[wicked sonnet]

yea sure

[wispy oasis]

who here has got a 4070 (not super/ti)?

[mortal delta]

I do not, sadly.

[wispy oasis]

i have one on the way but im eager to know how many sdxl images per second

[mortal delta]

you lucky i guess, that must have been pricey.

[wispy oasis]

about $550 usd + tax

[mortal delta]

yup, thats pricey.

[wispy oasis]

I need to generate several sets of 20k+ images so I need something that can do it fast. Stability api would take 12-24 hours I beleive for about 10k images so I thought its time to invest in a gpu

[mortal delta]

that is alot of images, may i ask what are you going to do with those images or is it secret?

[wispy oasis]

they're going to be used for NFTs using a new hybrid protocol that combines tokens and NFTs together creating fractionalized nfts

[mortal delta]

that sounds intersting, im not really into that kind of thing but i wish you luck.

[wispy oasis]

ty ty

[mortal delta]

i personaly would like to someday make some form of creative media with ai tools, but im overwhelmed to start.

[wispy oasis]

I think a 4070 can do 10k images in 6-8 hours or quicker. hopefully.

[wispy oasis]

thats fair you'll come up with something eventually

[mortal delta]

i sure do hope so, also i hope you have fun with your new graphics card.

[wispy oasis]

i will

i dont play games so

will be using stabe a lot lol

[mortal delta]

i play games here and there but not a whole lot.

[pine fiber]

EMAD RELEASE

it

[shadow dawn]

👌

[mortal delta]

hey guys, i finally got comfyui to work on my amd pc without crashing the system.

it took me 4 days

[cosmic loom]

didn't come here for a will, where did i go to prompt something plz ?

[lone hawk]

Yeah)) not gonna happen - he's waiting for that sweet deal from MS or some other investor with deep money pockets.

[pine fiber]

what deal man

its open source

[lone hawk]

not with SD3, you'll see soon))

[pine fiber]

bro knows something we dont

I know emad was hyping up soemething else that wasnt sd3

[karmic cedar]

https://youtu.be/dgTBScZOpT8?si=52DOMDDiS3hMdAK1 testing out SUPIR as a batch upscaler.

[winter pike]

it will be open source, but if its not im sure its gonna be leaked like SD1 and SDXL

[mortal delta]

dose anyone know of any projects i can do with ai tools such a stable diffusion or some form of chatbot? because everything i want to do seems daunting.

[karmic cedar]

Chatbot UI could be a good first step

[mortal delta]

i guess so.

[karmic cedar]

Have you checked that one out?

[mortal delta]

i somehow havent

[karmic cedar]

There’s probably a better option out there, but here’s an official link https://www.chatbotui.com

[hexed chasm]

What do you want to do?

[mortal delta]

but itint that kind are-iventing the wheel?

[mortal delta]

i want to make video games, mostly.

[karmic cedar]

Well, DeepMind’s latest abilities may yet become publicly accessible…

lol

[mortal delta]

oh interesting

[hexed chasm]

I was wondering if I could make animated cartoons - like take a comicbook panel and animate it

[karmic cedar]

Try to break your idea down into tasks, based on what you can accomplish on a step-by-step basis. Right now the approach to AI seems to be a bit like exploring an Adobe interface, where you’ve got various tools for various intentions

That’s my two dollars lol

[wicked sonnet]

SDXL turbo sounds so easy to replicate but is it? What I've heard it's "just" teaching a model to replicate the teacher more efficiently?

[hexed chasm]

Any idea how I might be able to take a panel from a comicbook and animate it?

[karmic cedar]

I think of these models as giant structured fractals that, when trained on themselves, are going to lose their complexity over time.

[hexed chasm]

I saw there's a Stable Video Diffusion, that accepts an image as input

But I don't know how good it is

[karmic cedar]

SVD is really cool—img2vid models in general have come so far, so fast.

If you’ve seen any of OpenAI’s Sora footage, then you’re automatically biased against anything else right now lol

[hexed chasm]

Well, can anything from Stability AI compare to Sora?

[karmic cedar]

Not at all, at present.

[wicked sonnet]

Well how come there's only one turbo model? I don't quite get the scope

[karmic cedar]

Ehh, not sure. That’s dependent on their logic

[hexed chasm]

Well, I figured that if I'm just trying to animate a panel from a comicbook, that this might be less demanding/difficult

I looked at examples in the #stable-video-diffusion chatroom, and they're most photoreal and don't look too good

[karmic cedar]

Carl Sagan had this quote about baking an apple pie from scratch, and how you first have to invent the universe. First things first—what’s the very first step you’d want to consider?

[wicked sonnet]

Well you sure you got the hardware to animate pictures?

[hexed chasm]

Well, if I have an Nvidia GPU, and I just want to animate a short clip -- just as a demonstration -- then isn't that alright?

[karmic cedar]

So for example—you have the idea, so now it’s time to get visuals from that idea. You could generate straight to video, but I would recommend a txt2img model first. And that’s where I imagine you’re at right now

[wicked sonnet]

Well anything you animate is going to use a lot of vram

or so I have experienced it

[karmic cedar]

So with that in mind, yeah—as Rem said, vram is a consideration. SDXL Turbo may have some LoRA files you could also use to boost the aesthetic look you’re going for, too—I haven’t really done my homework on that

so ideally, for this kind of thing— idea to text, text to image, then image to video.

[hexed chasm]

Hmm - LoRA to help with doing cartoons, eh? Are there any tutorials on LoRAs?
(I take Deep Learning in school, and I've been shown how to code LoRA directly in python - but I haven't really used any pre-canned GUIs, which people tell me are easier)

[karmic cedar]

I’m a DIY self-taught sort of person, so i tend to know what i know and that’s it lol

but let me pull up some other resources I like to use..

[hexed chasm]

Hey, thanks!

[karmic cedar]

https://civitai.com

Civit’s gonna be your model and LoRA hub.

[gaunt pulsar]

I wish there was a way to animate characters like in gacha games where they have an idle animation in a loop. I hope SVD is able to do that in the future, would be really cool

[hexed chasm]

okay!

[karmic cedar]

That’s a safe bet for you as well, because you’ll be able to filter the models down to specific bases that you’re after

[hexed chasm]

The artwork on this Civita page looks great - thanks - I just wish they'd show the prompts that generated them

[karmic cedar]

Often the model/lora pages will include descriptions of example prompts, but luckily those examples are pretty well indicative of what you can expect to see if you set them up properly.

SDXL is a more resource-intensive model, but the LoRA files spun off of that are often really fun. Like this one https://civitai.com/models/120096?modelVersionId=135931

[knotty turtle]

Is there an SDXL ControlNet that will transfer artistic style, but not color?
I'm looking for something like the old Shuffle model for SD 1.5. That model was amazing.
I've been experimenting w/ IP-Adapters, but they are transferring color as well.

[karmic cedar]

Good question—that I haven’t looked into lately. I know ControlNets are maturing faster than ever, and I’m just waiting for the next SD3 / Cascade generation of ControlNets to mature before I do another deep dive.

[hexed chasm]

gee, that looks like an 8-bit or 16-bit icon

[hexed chasm]

What is a ControlNet and what does it do?

[karmic cedar]

Also recommended, if you want to get your feet wet in a new model or LoRA—https://replicate.com. Replicate makes it insaaaanely easy to try new things when they drop, because they host all of the inference in these simple little pages.

But, you’re ultimately paying for those generations—so it’s only good to test with, and not to rely on.

[karmic cedar]

If a model is a big blob of code, then a controlnet is like structural scaffolding. When an input gets processed, the controlnets modulate how that input is diffused

[hexed chasm]

So the ControlNet allows you to have more precise control over the result

[karmic cedar]

Correct.

[hexed chasm]

How do you implement a ControlNet?

[karmic cedar]

And they allow you to isolate certain aspects or qualities to work with.

[hexed chasm]

If you want to make use of this ControlNet thing - then how do you do it?

[karmic cedar]

ControlNets are a part of the diffusion process. Think of them as like gears in a big machine.

And now we’re getting into…ComfyUI!

[coarse hollow]

yo guys stargin usnig ComfyUI today, when i've start the promt the consolle give me this error: "Prompt outputs failed validation
LoraLoader:

• Required input is missing: lora_name". Sorry for the noob question

[karmic cedar]

Sounds like you just need to refresh your interface and then make sure the LoRA file is selected in the module

[hexed chasm]

How do people use ControlNets? Is there some setting? Some extra piece of code you have to add on or install?

[karmic cedar]

It’s part of the workflow you’ll use when you generate images—but first things first, do you intend to generate them locally using your own GPU, or a cloud?

[hexed chasm]

Well, either one - cloud or local GPU (I only have access to local GPU part-time)

[karmic cedar]

Gotcha! Well, for now it might be easier to get acquainted with a cloud interface for ComfyUI.

That’s what you’ll use to get a more visual sense of what everyone is talking about.

[hexed chasm]

Okay, how could I find that? Would love to check that out

[karmic cedar]

one moment! 😄

Actually this would be a great one to crowdsource—does anyone have a good recommendation for cloud ComfyUI?

you could always try https://comfy.icu

[hexed chasm]

ah thanks - looking at it right now

[karmic cedar]

cool. big thing for most folks is the intimidation factor from the interface—just understand that it’s a bit like a coffee machine or a motherboard, you’ve got different components each doing different things and you’ll learn about them in time as they become relevant.

[hexed chasm]

any free ones? like on HuggingFace maybe?

okay, I found a Google Colab that does ComfyUI

So once we have ComfyUI up, then what do we do from there?

I see a workflow thing on my screen, but now I don't know what to do from here

[hexed chasm]

how do I make sense of this workflow thing?

[true canopy]

i have a serious lack of understanding how any of this ai drawing works, but im trying to learn

are the models/checkpoints on civitai made with stable diffusion as a base or whats going on?

[hexed chasm]

I haven't done SD for a couple of years, but I'm trying to get back into it

[true canopy]

actually i see now that it says "base model as 1.5" on civitai

[pseudo jetty]

Okay, so idk if anyone responded to you yet but this is the best i can explain it. The thing im talking about A1111 is a WebUI that allows us to generate images using whats called Models from Stable Diffusion. what you are talking about are Stable Diffusion Models and there are several different types of models that are used for different things. SD 1.5 is one of the most popular models. Simply because its capable of generating anime characters and NSFW type art a lot better than its successors. then you have SD 2.0 tbh i dont know much about 2.0 other than its better at real people than 1.5. SDXL is the newest model and its capable of better prompt recognition and better lora training, and overall higher resolutions than SD 1.5/2.0 Each one has its one purpose and strong points and it honestly depends on what you want to do

but the 1.8 and the 1.7 that i was mentioning are part of A1111 which like i said is simply the UI that is used to set your prompts and settings to generate images

[winter pike]

yes, theyre modified models with sd1.5 as a base

[true canopy]

thank you

[hexed chasm]

"SD 1.5 is one of the most popular models. Simply because its capable of generating anime characters and NSFW type art a lot better than its successors."
Ahh - really? 1.5 is better for cartoons and anime? Why is that?
And what are the drawbacks?

[stable zephyr]

sd1.5 had a company that had their finetune leaked

[true canopy]

apparently sd 3 is in some early release

[stable zephyr]

and they had put a lot of training into anime styles

[hexed chasm]

"SDXL is the newest model and its capable of better prompt recognition and better lora training, and overall higher resolutions than SD 1.5/2.0"

Okay, so I was thinking that LoRA could be another useful approach - but I find SDXL out of the box seems to give me the worst image quality results

[stable zephyr]

base 1.5 is pretty garbage tbh; but people have trained it extensively since that

[hexed chasm]

Where'd you hear this? I registered myself on the waiting list, but haven't come across anything anywhere

[pseudo jetty]

i think its because of the database they used to train SD 1.5 so SD 1.5 was trained on a closed database called CLIP, this database while it isnt public was most likely trained on anime and NSFW, however SD 2.0 is trained on OpenCLIP which was a open source and publicly available database that had NSFW images filtered out hence why it isnt good at NSFW

[hexed chasm]

wish I could find a rating site for SD versions, including the fine-tunes

[true canopy]

i should not have used the word "release" i simple saw it on the website
https://stability.ai/news/stable-diffusion-3

i have no idea about any release date , sorry

i should have said, more like, preview

[stable zephyr]

civit.ai has a very active community

[hexed chasm]

But could LoRA give me good results on doing a cartoon character, if I trained on specific images?

[pseudo jetty]

so loras a seperate model that are used on top of whats called a checkpoint, to produce specific results, SDXL would be considered the Checkpoint

[stable zephyr]

CLIP is the text encoder

[hexed chasm]

Gotcha - Checkpoint means "official release" - and LoRAs are fine-tunes (customizations) of those

[pseudo jetty]

yes thats correct my bad

[hexed chasm]

Yes, I'd read that CLIP is a pre-trained model within SD, that was trained on text-image pairs

[stable zephyr]

sdxl has two text encoders

[pseudo jetty]

CLIP was trained on a private dataset, where as OpenCLIP was trained on a different dataset

[stable zephyr]

CLIP = ViT-L/14 https://huggingface.co/openai/clip-vit-large-patch14

[silver reef]

where can i have help ?

[pseudo jetty]

but yes, lora is what you would use if you want to generate an image of a very specific thing.

[hexed chasm]

Speaking of fine-tunes/LoRAs - how can I find the best fine-tune available for cartoons & anime?

[pseudo jetty]

but you would use that on top of a checkpoint such as SD 1.5 or SDXL

[pseudo jetty]

Civitai has a very large collection of publicly available loras

[stable zephyr]

all stable diffusions were trained on a 'fixed clip'; whereas subsequent finetunes also train CLIP

[true canopy]

im getting pretty bad results, but i cant tell if its because people are doing very high resolutions on civitai, or if im doing something else wrong, especially eyes come out bad

[pseudo jetty]

are you using highres fix

and what are you trying to generate? and on what SD 1.5 or SDXL

[true canopy]

im using a upscale model, but ive seen highres fix a lot on civitai, is it just a upscale model?

[pseudo jetty]

no

[true canopy]

oh its not? then i am indeed doing something wrong

[pseudo jetty]

so do you use a1111 or comfyui

[stable zephyr]

it's a process of upscaling and doing further processing

[true canopy]

comfyui

[pseudo jetty]

oh im not familar with that, someone who uses comfy ui would be better to tell you how to do it

[stable zephyr]

If you take and drag one of the images from my model into comfyui, it will show you the workflow I used: https://civitai.com/models/239909/darkclip-25d

[true canopy]

lol im getting so many errors when i try to do that

since im on just base comfy

[hexed chasm]

Does SD XL allow Negative Prompts?

I tried them, and they seem to have no effect

[stable zephyr]

if you go in to options there is a tool to auto-install all the dependencies

[true canopy]

rly? i had no idea

[radiant tusk]

what's the best face detection model for adetailer?

sometimes there will be one face and it would detect like 7

[true canopy]

i, dont see that option in options

maybe its because i dont have the manager?

[stable zephyr]

oh sorry, yes, it's in manager

[true canopy]

alright, now i just gotta figure out how to get that manager lol

[stable zephyr]

https://github.com/ltdrdata/ComfyUI-Manager?tab=readme-ov-file#installationmethod1-general-installation-method-comfyui-manager-only

you just git clone it into the extension (custom_nodes) directory

[true canopy]

lol i cloned it into the wrong dir, but i got it now

[fervent thunder]

Is the seed related to the model

[true canopy]

this stuff is incredible

[fervent thunder]

or can I use the same seed + prompt with multiple models to produce the same image in various artstyles

[true canopy]

it seems im getting an error using this workflow, seems its related to my amd gpu, but thanks for showing

[shell tendon]

GitHub just went down I think

[stable zephyr]

when you installed comfy, did you make sure you were installing the rocm version of torch?

[true canopy]

ehm, good question, i simply did pip install torch, i assume i didnt get the rocm

[stable zephyr]

welcome to the ecosystem heh

[true canopy]

yes...thank you lol

i guess...ill try to uninstall and redo it

[pseudo jetty]

as far as i know they do allow it, but they arent useful nor necessary. The SD team attempted to make it unnecessary to use negative prompts, so i dont think they work.

[pseudo jetty]

so what are you trying to generate if you dont mind me asking, if its cartoons, and anime, honestly id recommened SD 1.5, or a Fine tuned SDXL model for Anime

also when using SDXL and Loras, you will need whats called an LoRa XL its a specific type of lora for SDXL

[hexed chasm]

Strange they don't see negative prompts as necessary, when SD XL has inferior image output

[pseudo jetty]

SDXL is supposed to have a much better image output

but again it depends on what you are trying to achive

SDXL has a lot more fine tuning espcially when it comes to the human body

SDXL i believe is also capable of generating text

[patent scroll]

Hi. Previously I worked on some scenes on local Automatic1111 SD 1.6, now I want to continue on Automatic1111 SD 1.7 but they don't come out the same.
I guess that's normal? Can I make SD 1.7 process those few images as if it was 1.6?
Otherwise it's the same model and parameters (from pnginfo of the last renders).

[pseudo jetty]

as far as i know upgrading to A1111 1.7 from 1.6 shouldnt effect image generation?

[patent scroll]

Could it be newer pytorch causing it?

Or using ZLUDA?

[pseudo jetty]

as far as i know 1.7 didnt update torch, but update 1.8 did

im not sure what ZLUDA is so im not sure on that part

[patent scroll]

ZLUDA translates CUDA to use on the AMD GPU with near native performance

[pseudo jetty]

you know what, i know 1.7 added HyperTile Support, that effects image generation

also i know some settings in 1.6 were broken and werent doing what they should have done, so when you updated to 1.7 and continued to use those settings, they will now behave differently

[patent scroll]

Thankfully I installed 1.7 separately

[pseudo jetty]

i mean you can also just revert, you dont need to keep them seperate

i had updated to 1.8 and a few things were broken, so i just downgraded back to 1.7

[patent scroll]

It is simpler and safer to keep them separate. They share the models through hardlinks anyway.

[pseudo jetty]

i mean ive never noticed anything different but whatever works best for you

[patent scroll]

Likewise.

[pseudo jetty]

but what do you mean its safer to keep them seperate?

[patent scroll]

Just in case one breaks (or I break it) I have the other one immediately available.

[pseudo jetty]

ah, that doesnt matter much to me. it only takes a few seconds for me to reinstall if it breaks

[patent scroll]

And some things may work (better) with one and not with the other.

The 1.6 torch is 2.0.0+cpu and the 1.7 torch is 2.2.0+cu118 with CUDA 11.8

[pseudo jetty]

ahh, okay. see i didnt see that in the notes for the offical release

[patent scroll]

Please mind I keep talking about the Automatic1111 version

[pseudo jetty]

yes i am aware

[lapis urchin]

if i have multiple instances of the webUI running and change the output folder for one of the instances it wont change it for all instances correct?

[true canopy]

it seems i gotta give up on the highres fix, its simply to taxing on my rid, one image is like 30min

[pseudo jetty]

how many highres steps are you doing

[patent scroll]

CUDA wasn't on the first one because it was just DirectML for AMD, and the second one has CUDA because of the AMD ZLUDA

[true canopy]

just one lol, its just img to img seems like it needs better gpu i guess

[lone hawk]

Lmao what's your gpu?

[pseudo jetty]

thats what i was gonna ask

[true canopy]

its an old amd gpu, 5700 xt

[pseudo jetty]

dog thats newer than man

[true canopy]

or im doing something wrong

[patent scroll]

Could lowvram and medvram arguments make a difference in generating images?

[pseudo jetty]

they say it doesnt but i think it does

[true canopy]

prompt is telling me im getting 43s/it, no idea if that fast or slow, im assuming slow tho

[patent scroll]

So I had lowvram on 1.6 and medvram on 1.7. I will go test it now.

[pseudo jetty]

what are you highres fix settings

brb ill respond in a minute

[true canopy]

i just dragged an image from civai, i barely know how these things work yet

[pseudo jetty]

im downloading comfyui now let me check

send me a ss of your workflow

[lone hawk]

Why are u using comfy? This is for experts, use latest Automatic 1111 1.8

[pseudo jetty]

agree, comfyui is for more advance workflows

[true canopy]

i want to learn, its my first day

it seems easier to understand with the workflows

[lone hawk]

no, easier to learn A1111 and then comfy

[true canopy]

and it seems like u have a lot more control of what u want to do

[shell tendon]

other way around IMO

i had no idea wtf was going on until i switched to comfy

[lapis urchin]

i have a question; with all the images i see here they seem way better than mine, is it just mostly generating tons of images and fine tuning the prompts for a really long time?

[honest spear]

Workflows are better if:

1. you want to learn how stable diffusion really works (it's just 5 6 nodes to get it working)
2. you want to more freedom or specific extensions
   A1111 is not bad for beginners but Comfy isn't that difficult

[lone hawk]

true) also you might want to get an Nvidia gpu, good luck!

[true canopy]

ive had a lot of fun today on comfy, u can def just run it immidiatly with just a few nodes, but i do agree that if u want great results, the nodes get a lot more complicated

but that goes with anything in life, if u want great results u gotta put the hours in

i doubht the people with great generated images, just moved some sliders on a1111 and they popped out

[patent scroll]

@pseudo jetty I tested all, lowvram, medvram, medvram-sdxl, and no vram arguments. No difference between any of them. They all rendered identically different to 1.6

[honest spear]

mmh it depends, sometimes the model is really really good at doing something, and highres fix can do wonders in some cases

[true canopy]

well, either way my gpu cant handle it sadly

[patent scroll]

I currently test some realistic model in A1111 and short prompt of around 10 tokens, and 5 negative tokens, generates a batch of 100 images with at least 5 in it amazing just needing some minor inpainting

Sliders have almost nothing to do there, the tokens matter.

[lone hawk]

A1111 ver. 1.8 add something new called Soft inpainting

[clear oyster]

for jugernaut and dreamshapers everything is coming out like it needs more than 4 steps

[pearl ocean]

[fresh igloo]

Being slightly overwhelmed, I was curious if anyone had some advice on getting started? I have a pretty good PC so I'd like to use local resources as much as possible

Start with Automatic?

[hallow sable]

lol 4 steps is too low

[shell tendon]

start with forge and/or comfyui

[rotund forge]

Agreed, Forge is a fork of A1111 by the guys who develop Controlnet: https://github.com/lllyasviel/stable-diffusion-webui-forge

More optimized than A1111 at the moment - think there's a one click installer if you're not up on git/python/etc

[fresh igloo]

Awesome, I'll give that a go.

[rotund forge]

They also do "Fooocus" which is intended for people who don't know anything techy/SD settings and just want a Midjourney experience.....most of the stylistic stuff (steps/cfg/etc) don't need to be understood: https://github.com/lllyasviel/Fooocus

Comfy is likely too much unless you're a developer type or at least are planning to put in time watching youtube tutorials to get rolling......

[wispy oasis]

is stablility api down

[hallow sable]

do forgeee work with dreambooth?

[fresh igloo]

Not a developer, but an industrial automation engineer - so enough to be dangerous

[pearl ocean]

Can’t wait till SD can do like 1 min videos or longer

[hexed chasm]

Are you speaking of Stable Diffusion or Stable Video Diffusion?

Can regular Stable Diffusion do videos?

[sterile raven]

They said they are testing SD3 for performance and safety. What kinda safety are they testing for? That they didnt accidently create a malevilant AGI???

[prisma horizon]

maybe they are testing that inadvertently loops and makes pc explode

[thin notch]

hi guys can u tell me that deforum extension work? bcs i got ModuleNotFoundError: No module named 'deforum_api_models'?

[prisma horizon]

you need to download the model for the extension

unless the extension can't find the model

[thin notch]

okey thanks i will try it from where i can download it ?

[radiant tusk]

what upscalers do you guys use?

[meager trench]

I donloaded a bunch or Loras and Lycoris, but forgot which safetensor file is a Lora and which a Lycoris, is there an easy way to select them and move them into the right folder

[lapis urchin]

Civitai

[prisma horizon]

just place all of them in lora directory XD

[meager trench]

Hehe, is there a log setting that I can see with an error message so I can pick them out?

[prisma horizon]

even when in lora directory, they should work the same

[meager trench]

ah great, phew 😄

Thanks 🙂

I was on Civitai and thought, wow that looks cool, and that and that and that and wow look at this....

[mortal delta]

does this server have a tech support channel that i missed?

[prisma horizon]

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

[mortal delta]

i feel like i always lose track of where channels are.

and i have to always ask where they are.

when in reality i somehow skip over it.

[prisma horizon]

happens to me too

[mortal delta]

good im not alone on that.

anyways i made a post in the tech support channel, so hopefully i can get something out of it.

[warm junco]

Yes thats done via the x/y/z script

[meager trench]

Thanks for mentioning automatic1111 1.8 above... updating 🙂

[warm junco]

You need to select steps for x, then set the value to 20-30 so it will do 10 images from 20 steps up to 30 for example

You can also specify like, 15, 20, 25, 30

Yes I updated it a day ago