💬｜general-chat

[trail lion]

there's really no model-based concept of time coherence, so the best you can do with current tools is force it with controlnet or whatever

[loud solar]

I remember that A.I. using Stock Videos to create a small film ... but for sure the content has been pre-analysed ...

[trail lion]

I see posts in the stable-video-diffusion channel here, or in the SD discord on reddit, I have yet to see something that makes me thing the tech is much further along

[loud solar]

For sure you can analyse the content of 30 videos, trying to get the most important points and cut them to a new video.

But I don't think theres a solution to do it out of the box yet

I think it's been https://invideo.io/

[trail lion]

that's an interesting txt2video implementation. I just watched a youtube demo of it. but I doubt that's what most people here are trying to do

[loud solar]

Yes ... it won't help him/her. I just wanted to say the technology already kinda exist. But not yet made to analyse your videos and put them together again. I've tried that A.I. in a pretty early stadium ... it worked quite good.

I need to sleep. Good nite!

[atomic summit]

Hi, I am getting "Torch is not able to use GPU; add --skip-torch-cuda-test to COMMANDLINE_ARGS variable to disable this check" this error and I suppose it is related to running pytorch with gpu. I am very lost at this point. any suggestions?

[trail lion]

probably an issue with your python environment, did you check the https://discord.com/channels/1002292111942635562/1002602742667280404 channel?

[atomic summit]

sorry I didnt see that. I'll write there then

[frigid sinew]

Why is my stable diffusion output so bad

like the same prompt gives me way better results on Chat GPT

Do I need to use keywords?

Like I will type "Peter Griffin laying down with his feet kicked up and in close eye view, the feet are particularly stinky and putrid with visible signs of dirt and grime built up"

And the output in Chat gpt is wayyy better

but the issue is that sometimes it gets blocked by the filer

[charred mesa]

finetuned SD3 models will possibly be able to do that

can't promise anything

[narrow kernel]

What does "stinky and putrid" look like? What does "feet kicked up" look like? What do" visible signs of dirt" look like? Why not just say dirty or grimy feet?

Starting to see where I'm going with this? You have to ask yourself these kinds of questions when making a prompt. GPT has had a ton more time and users to DPO things around people using bad or vague prompting.

[glossy jewel]

This happened to me the other day when I did my install. The problem is that the version of torch that got installed isn't the a version that uses cuda. The install script gets the CPU version by default. If you know how to use Python, just uninstall the version of torch you have an pip install the one using cuda.

[young surge]

I did find another fix for it.

Open a new tab, (On the same window, and go to new tab), wait a few sec
Click on the WebUI cmd window. press enter a few times. Wait for new activity to occour. Press enter a few times again, and that should have cleared up the gen xD

[frail linden]

I keep trying to subscribe to Stable Diffusion but the page never goes anywhere. Is anyone else having this issue?

[young surge]

Try a new browser

[frail linden]

I tried Brave and the Microsoft one. Same issue.

Edge

I select what subscription I want, it goes to a new page, but that page never loads up.

[hasty geyser]

Hi guys
Is Easy Diffusion as good as Stable Diffusion ?

[frail linden]

Is StableDiffusionWeb subscription the same as StableDiffusionAi.ai ?

StableDiffusionAi won't take me to the subscription page but StableDiffusionWeb will.

just tried on Opera with no luck.

I'll try again tomorrow.

[rare parrot]

That sounds like a scam

The company name is Stability AI

[frail linden]

It's definitely confusing.

[rare parrot]

easy diffusion sounds like some program to run stable diffusion or some scam that will run stable diffusion for you on some remote PC while charging you 10x the industry standard for that amount of compute

@frail linden https://stability.ai/stable-assistant

[frail linden]

I thought this was Stable Diffusion. Am I wrong? https://www.stablediffusionai.ai/

[rare parrot]

Yes

[frail linden]

What website is that then?

[rare parrot]

I just linked you

[frail linden]

It generates AI art.

[frigid sinew]

Lets ignore the obvious fetish undertones xD

[rare parrot]

That subscription gets you access to their Discord bot

And it lets you use Stable Diffusion 3

That's the only place where you can use it too

[rare parrot]

Yeah that's basically a scam

That link you got

They named a website to confuse people

[frail linden]

so what is the correct website?

[rare parrot]

Here is the official one

Also you can use a discord bot in this server for free to use Pony model, which is amazing tho less good than Stable Diffusion 3 on that website

But Stable Diffusion 3 can't do nsfw I believe and probably doesn't know as many fictional characters

Since it's an official corporate product basically

[frail linden]

boy, this is confusing. I liked the site I linked since it was simple enough to just put in the text and it made stuff.

[rare parrot]

Yeah there are a lot of websites like that

And you can even download a program for that for free

So most unofficial websites will basically try to charge you extra money if you don't know better

Avoid them if you can

[frail linden]

so that website was charging me to use their discord bot?

which is free

[rare parrot]

This one

Wait

OH

Lmao

I thought this is a different server

For a sec

I though this was Pony Diffusion discord

one sec

[frail linden]

hmm

[rare parrot]

Yeah I have a tall monitor, text box is on the bottom

I forgot I changed which Discord I have opened

Anyways

You can use google colab I believe to run Stable Diffusion for free, or basically any website that lets you run your own programs remotely (since all models based on stable diffusion are free software)

And if your computer is good enough you can run that stuff yourself

For free, with a lot of extensions that can do nice things

Etc

[trail lion]

you cant do it on colab free, it's against google's EULA. You can do it on their pro tier, but there are probably better options like rundiffusion

[rare parrot]

Ah, makes sense

Well, local is the best for most purposes but it requires having what would be considered a gaming computer

[frail linden]

I was using this site https://www.stablediffusionai.ai/ but it asks me to subscribe since you have limited credits plus watermark.

I just want to put my text in and get images. I don't need to re-engineer and set it up on my computer.

[rare parrot]

Do you care about it being free in the end or not really?

[trail lion]

then you'll have to buy a service to do that

[rare parrot]

Yeah with few exceptions

[trail lion]

there are many out there, like the one you found

[rare parrot]

https://stability.ai/ is the official website here, you can find subscription options and they have a free trial

It'll let you access a Discord bot in this server here

Which lets you use Stable Diffusion 3

Which is pretty much the best model right now

[frail linden]

so it does same of what that website does?

[hasty hornet]

anyone has background removal + background inpaint / change workflow for comfy?

[rare parrot]

More-less but you use a Discord bot you text instead of a text field, that's the only difference, plus it'll work better

There is also https://www.bing.com/images/create tho I don't know how much the images cost there and if there is a free trial nor how long

[frail linden]

so discord bot is typing text into a chat channel? Doesn't that mean anyone can see your image?

[trail lion]

if you need privacy, then you want the local 're-engineer it' option

[rare parrot]

I head rumors they'll allow to use the bot outside of publically in chat here in some time

Anyways setting it up locally isn't all that hard and you'll have no limits on how many images you can generate so that's why I prefer it personally

And the fact that nobody can see it is nice too

[trail lion]

tons of reasons it's better, you ultimately will want to save some portion of them, and if they are already local that's so much easier, plus loading them into controlnet, etc, I cant imagine doing all that on a website

just being able to generate images online, while convenient is also really short-sighted

[rare parrot]

Most websites don't have controlnet or anything like that

[frail linden]

if it is local, how does it scan the image library? Are you saying the image library has to be downloaded to your computer as well? That sounds like a big file.

[rare parrot]

No

The AI itself is like 3 to 6GB

And the programs to run it less than 2GB

[frail linden]

so it generates the images to your local computer

[rare parrot]

And as long as you have a decent graphics card you can generate unlimited images for free

Yea

No need for internet or anything

It's just a program like any other

[frail linden]

I have a Nvidia GeForce GTX 1070 with about 32 gb of ram for this computer. Is that enough?

[rare parrot]

More than enough

[low moon]

Everything is fine....

[frail linden]

It's not like I'm playing Star Citizen on ultra or anything...

[low moon]

Its just one model

and another one and another one

and another....

[rigid loom]

Hello guys

[hasty hornet]

ain't gonna be fast, but will work...

[low moon]

Patience is a virtue.

[rigid loom]

I'm tried installing comfyui, portable, etc and I just CAN'T make svd model work anywhere, I'm uninstalled and installed over 10 times and always some error, can anyone help me out giving me some sort of sure fire way to make this work?

[rare parrot]

It might take as little as 20-30 seconds with that hardware depending on what quality images you'll be generating

[rare parrot]

svd?

[rigid loom]

stable video diffusion model

I want to do txt2vid

[rare parrot]

Oh lol I'm staying away from txt2vid for now I think

[low moon]

video isnt there yet

come back in a year

[rigid loom]

How come the websites make it work then?

[rare parrot]

They have some proprietary models they don't publish

[rigid loom]

I keep getting python errors everywhere

Ahhhhhhhhhhhh

[low moon]

u trying to run collabs?

forget it

its a mess

[rigid loom]

is that why everything I run is wrenched?

[low moon]

local or uhm go... home?

[rare parrot]

Anyways with local models currently best you can do is basically applying a filter on a pre-existing image of having things move around like they're spasming

[rigid loom]

thats what im trying!!

I'm so frustrated

[rare parrot]

Define wrenched

[low moon]

I was too

[rigid loom]

wrenched as in, the tutorial guys make it work and I can't

[low moon]

learnign all this is a head ache

im barely getting into comfy

[rigid loom]

so generate pics then animate?

kinda lackluster for me tbh, I'm trying to work on content for video projection in places like clubs etc

[rare parrot]

Sorta but also not really

https://youtu.be/G7mihAy691g

This is more-less like best case ideal scenario

Of what local models can do for video

[rigid loom]

Thats what I want to do

now how do I do it?

[rare parrot]

It will look like a mess at a glance, like I'm not sure you're realizing how low the quality on those is when you look closely

Movements are never natural looking

[rigid loom]

so for big screen projections it's not gonna fly?

[rare parrot]

And the framerate is super low without some extra interpolation

Yeah

Oh and the videos you can generate can be max like 5 to 10 seconds long

[trail lion]

yah, the 'videos' for SVD are essentially like if you took an image and just slightly rotated it, while keeping the background fixed

[rigid loom]

that's fine, I can loop them and play around on the VJ table, y'know

[rare parrot]

Mm

[rigid loom]

but what about that corridor crew type shit people do?

[rare parrot]

Well, Comfy is probably your best bet

What errors are you getting?

I can probably help you with that

[rigid loom]

well I uninstalled it all in a fit of rage

and now want to reinstall stuff properly from scratch

even learned how to do shell commands and stuff

my brain hurts so much

[rare parrot]

I just do everything in shell hueh

Tho mine is way nicer to use

With stuff like autocomplete based on history and such

[rigid loom]

I tried through Anaconda, Pinokio, and most recently through Github clone repository

[trail lion]

the corridor crew stuff, it's a whole other thing, it opens up the possibility of what you can do to some extent, but the limits are still going to be something like 1-2s of video and there will be flickering

[rigid loom]

github's Comfyui doesn't even download the models / workflows

[rare parrot]

Workflows are just in a single image metadata

You can grab it form anywhere

[rigid loom]

I can deal with flickering, it's fine

[trail lion]

flickering because the current image diffusion models dont have a concept of time coherence

[rigid loom]

my boss thinks it's high tech shit, y'know

[rare parrot]

Yeah it's more of a tool for multiple jobs on purpose

[rigid loom]

can I share some stuff that he approved already?

[rare parrot]

Probably yea

[rigid loom]

I have to make like 30m total content

[rigid loom]

I see

[rigid loom]

as long as colors don't flicker, I'm good

or as long as he manages to make animals and stuff decently

[trail lion]

colors...not sure, it's more like objects will appear and disappear, etc, because despite all efforts to keep the image mostly the same from frame to frame, there will be differences, every image is unique in SD

[rigid loom]

#🌶｜off-topic message

here's one of my generated stuff

[rigid loom]

thats fine

#🌶｜off-topic message and here's another

[rare parrot]

I see

Not bad

[rigid loom]

thanks! Was hard

[trail lion]

hair is a good example, thing of a woman's long hair...think of generating the same image of a woman 10 times, her hair will be in a different spot every time, and when you animate it, the hair will flicker

[rare parrot]

Well, as an abstract AI animation

[rare parrot]

Depends

[rigid loom]

yeah don't need that much level of detail tbh

[rare parrot]

Not with new video models

But those have big limitaitons still

When it comes to image quality

[rigid loom]

it's more to set a vibe, with colours movements, sceneries

[rare parrot]

And what movements they can handle

I see yeah

[rigid loom]

people at a club drunk, at the end of the day lol

[rare parrot]

Anyways

[rigid loom]

the second video has some capcut editing ofc

[low moon]

you can try haiper

[rare parrot]

There is a very nice ComfyUI extension called Manager that lets you download some extensions and models and looks like it lets you download that one stable diffusion video model with one click once you have the Manager installed

I'm downloading right now

Going to test official ComfyUI workflow for it

[rigid loom]

Okay so from scratch what version of comfyui should I download and where should I install it?

I've had this stupid python environment error once I couldn't fix

[rare parrot]

https://github.com/comfyanonymous/ComfyUI

If you follow instructions here, you should be all good

The portable standalone build thing for windows

Unless you're on an AMD card

Tho I doubt that

I am tho personally

[rigid loom]

yeah nvdia here

ok downloading the zip, extracting, easy enough so far

[rare parrot]

Should be hopefully yeah

After that see if the program opens and such, then I'll link you that manager extension that makes downloading models and extensions easier

[rigid loom]

ran through the powershell commands first, right?
" Downloading https://download.pytorch.org/whl/nightly/cu121/torch-2.4.0.dev20240520%2Bcu121-cp311-cp311-win_amd64.whl (2481.0 MB)
--------------------------------------- 0.0/2.5 GB 11.7 MB/s eta 0:03:30"

or should I be doing this in the comfyui terminal?

[rare parrot]

Nah it doesn't matter I believe

Also there is no comfyui terminal

[rigid loom]

kk just to make sure

then..

what does this mean:

"Dependencies
Install the dependencies by opening your terminal inside the ComfyUI folder and:

pip install -r requirements.txt

After this you should have everything installed and can proceed to running ComfyUI."

[rare parrot]

It's the Windows terminal

In which you navigate to Comfy folder

[rigid loom]

okay damn

so on powershell

[rare parrot]

Just wanted to avoid confusion

[rigid loom]

how do I navigate there?

yes yes, thanks a lot man, I'm super noob at this

[rare parrot]

dir shows you all folders in your current folder

cd folder where you replace folder with actual folder name makes you go to that folder

cd .. makes you go to one folder back above the one you're in

[rigid loom]

okay let me try that out

[rare parrot]

I believe it puts you in C:/Users/your-user-name/ by default

[rigid loom]

yes comfyui launcher is there

but how is that related to the Comfyui master I just downloaded?

[rare parrot]

master is basically just the name of the software branch the comfy devs work on, like basically master is the up to date version of the code, then something like dev is usually where they work on in real time and once a feature is done they push the changes to master

So tl;dr doesn't matter

But what do you mean by that

You have two separate ComfyUI things?

[rigid loom]

because that's what I downloaded I think?

[rare parrot]

Basically use cd and dir to go inside the unzipped comfy folder

[rigid loom]

ok doing that

[rare parrot]

Also is this still running or not anymore?

[rigid loom]

just finished, it's updated

wanted to do the requirements now

[rare parrot]

I believe that was one of them so it probably just auto installed

[rigid loom]

trying out through what you taught me now

I see

omg I did it

[rare parrot]

pip install -r requirements.txt

Means:

pip = package (program/library) manager for python
install = telling it to install
-r = read list of stuff to install from a file
requirements.txt = name of text file that list is in

[rigid loom]

it's installing the requirements and I did it through CD 😄

[rigid loom]

dropping off from programming to pursuit arts was the biggest mistake of my life

first error message

"ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
torchaudio 2.2.0.dev20240520+cu121 requires torch==2.4.0.dev20240520+cu121, but you have torch 2.2.1 which is incompatible.
torchvision 0.19.0.dev20240520+cu121 requires torch==2.4.0.dev20240520+cu121, but you have torch 2.2.1 which is incompatible."

but I guess it installed everything else

[rare parrot]

Hm

Yea

Alright so basically:

ComfyUI zip thing I beleive comes with it's own install of Python I believe

And you might've installed Python on your PC earlier I believe

It shouldn't cause issues but if it will we'll take care of that

What matters is: Python has something called venv where basically you change where your cmd or powershell thigns Python is installed from the actual installation folder on your PC to some folder inside a program like ComfyUI

[rigid loom]

alright, I see, yeah tons of them installed and uninstalled, I tried using Automatic1111 but ran into that python issue cuz of versions

[rigid loom]

Oh, so thats what venv does

[rare parrot]

That lets you have different versions of Python and libraries it uses for different programs on your comptuer without conflicts

Yea

Basically separate per-program Python install

Using that I can help you fix any dependency issues later on

For now, I understand it said everything is installed?

[rigid loom]

yep, executing run.shell now

[rare parrot]

Nice

[rigid loom]

#🌶｜off-topic message this is what we got now

I have some workflows but when I drag them it seems like it doesn't really install the custom nodes

[rare parrot]

Not sure if you saw what I replied but anyways

That "Launcher" thing

Seems to be some other program

First time I see it honestly

Did you download that one in the past?

[rigid loom]

Yeah

I guess the one you sent me is still in its own thing inside the folder

[rare parrot]

Ah

Yeah let's try using that one

cd into that folder if you manage

And also

We should probably move this to private chat or #🤝｜tech-support since we're kinda spamming this channel

Which one do you prefer?

[rigid loom]

Let's do DM

[rare parrot]

Alright

[cursive kindle]

Is this stable ronaldos discord server

[trail lion]

no idea who renaldo is

[nova zodiac]

given his haircut around 2001-2002, I don't think Ronaldo has been stable for a while...

[ember imp]

https://tenor.com/view/trollszn123-ronaldo-gif-18268194

[rigid loom]

Why, yes I am Portuguese

[ebon dagger]

Is this discord server specific to stability.ai or stable diffusion in general?

[frail compass]

How can I generate a picture of AI

[desert dagger]

stable diffusion in general

[frail compass]

?

I want an AI picture but where can I get it

[ebon dagger]

Ok thank you

[sly eagle]

Hello!, I've been watching many posts on social media about ai artists copying a certain generated image. How do I know if a image I generated is not a copy?
what would be good for more original pictures, checkpoint or lora?

[nova zodiac]

Ai copying another ai image, or ai copying an image from a person??

[nova zodiac]

A lora is a mod for a checkpoint, doesnt work on its own

[stark aspen]

My company upgraded me from a 16" M2 Max to a 14" M3 Max (unbinned 16/40 cores) and my SD generations (same pictures for testing) have not improved (they're the same or slower). I am using torch==2.1.2 nightly with MPS support and have applied all the recommended tricks on both machines, in any case: they're exactly the same performance:

(((Drawn by Alena Aenami))), Intricate details in serene landscapes, peaceful, Alena Aenami's artistry. Ethereal sky, gentle cloud formations, soothing light rays dance across a breathtaking tranquil beach scene; soft sand meets the ocean waves beneath an azure blue horizon. Nature's canvas: smooth, flowing lines, soft colors blend harmoniously to evoke emotions. Serenity permeates this enchanting setting.
Negative prompt: human,face,person,people,boy,girl,woman,man,
signature,logo,text,letters,username
Steps: 30, Sampler: DPM++ SDE, Schedule type: Karras, CFG scale: 7, Seed: 2545687968, Size: 640x1536, Model hash: d91d35736d, Model: juggernautXL_juggernautX, Clip skip: 2, Version: v1.9.3

Time taken: 1 min. 37.8 sec.

M2 Max -> 3.4 it/s
M3 Max -> 3.35 it/s

About 90s every time. I'm seeing this is a pretty average score for A1111 so I'm not worried about that, just wondering if anyone has had the same experience, or if the difference is actually the 16" -> 14" thermals downgrade (smaller chasis=hotter=throttle?)

Anyone that can test for me on their 16" M3 Max (16/40)?

[zealous wraith]

https://tenor.com/view/welcome-gif-26939290

[dim portal]

https://tenor.com/view/alvin-and-the-chipmunks-alvin-whoops-my-bad-oops-gif-15512287650458333097

[fervent thunder]

hi guys I am running into a problem

need help

trying to train a lora

and it gives me this: PermissionError: [Errno 13] Permission denied: '/config_dreambooth-20240521-111115.toml

[unkempt hatch]

so lightning and hyper models are models which have had 'junk' removed (distilled) and are therefor faster, require less steps.

If I'm creating a mix, and include some lightning/hyper models, then my mix would be somewhere in between? So say it has 6pony, 4sdxl, and 2 lightning, I'll get something slightly more efficient, requiring maybe 20 steps instead of 30, right?

Jujst trying to undersdtand it better

[teal pagoda]

I don't advise you to do that

mixing checkpoints with differents architectures

because in most of the cases, the resulted model will generate black/one color images with a very low quality.

[pearl ocean]

People these days could easily get by with the iPad 10

[crimson tendon]

Hey guys, I've been having a bit of trouble getting SD to work with my AMD gpu. Linux main system ryzen 5 3600, RX6700XT. The system should be fine running it I just need a bit of setup help. If there are any wizards that can help me work some magic that would be great

[teal pagoda]

the best "wizard" will cast a spell to make money appear in front of you, so you can buy a Nvidia 4000 series GPU for running Stable Diffusion

[crimson tendon]

haha, I am the joke yes.
Wanna help or keep laughing in CUDA?

[tall musk]

they both sound a bit on the slow side, are you using fp16 or fp32 ?
I'd expect approximately 2.1 s/i for the M2 (38 core ?) and 1.5s/i for the M3(40 core) under good thermal conditions using fp16.

[trail lion]

Go here https://discord.com/channels/1002292111942635562/1002602742667280404 find the pinned messages at the top for a guide

[lofty sphinx]

Bruh european ai act just got approved by law

[full lark]

What was that about? 👀 🥺

[lofty sphinx]

https://www.consilium.europa.eu/en/press/press-releases/2024/05/21/artificial-intelligence-ai-act-council-gives-final-green-light-to-the-first-worldwide-rules-on-ai/

L EU

basically u can get arrested if ur project doesn’t have a watermark atleast from what i understood

sad to see the country that has less development js starts regulating AI lmfao

[cedar salmon]

High-risk AI systems, as well as certain users of a high-risk AI system that are public entities will need to be registered in the EU database for high-risk AI systems Gonna have to register if you make memes

[full lark]

mfs EU sucks 😭

[cedar salmon]

i think most will stumble through legislation poorly

that would be normal

[lofty sphinx]

how am i supposed to put a watermark on the output of generated mc skins project i made if that would literally ruin the whole porpuse of the project

[cedar salmon]

you could register as a non watermarked photo is high risk AI

everything will be fine

[granite rover]

Hope i didn t break any rule by posting this

[finite rain]

[astral lotus]

I love how you got 3 mistakes in your phrase and it's (edited) already.

Really makes us want to give you money 😅

Let me help you:

I hope I didn't break any rules by posting this

[granite rover]

It s human to err. I apologize

We are a newly established company. And we are still in the market testing phases. At the moment we are working on the website.

[astral lotus]

You established a company BEFORE a market research and a website?

[granite rover]

yes

We are currently listed on vast.ai.

[astral lotus]

No offense, but I would take a few Business classes before doing this project.

A company establishement should usualy be preceded by a market study 😛 To see if you actually have market place to win

No market study leads to 4 corner stores at the same corner

[granite rover]

I think we did our calculations quite well. We have already purchased the equipment. So I don't see what's wrong with that. It was simply a re-profiling.

[astral lotus]

Anyhow you do you, but remember that you now represent a company, so use an auto corrector. There's no shame in that 😛

[granite rover]

Right, I apologize again for that error.

[royal sentinel]

ive had success training LORAs in Kohya ss.. but have recently been testing out my ability to train full checkpoints with Dreambooth... to no avail.

research into whats going wrong is also leading me to question... is there an advantage to training the full checkpoints or is LORA the way forward?

I was thinking that by training some style concepts into a dreambooth i would have better luck combining that model with addional loras at generation stage, where combining multiple loras with signficiant influence can be tricky....

anyone have input as to lora vs dreambooth training full models??

[teal pagoda]

I'm not laughing, I'm an AMD user myself saving for the 5000 series. I'm using cloud Nvidia GPUs for now, because AMD is too bad at this

[thin turret]

how do i access the prompting channels?

[sly eagle]

how do you download fast some checkpoints from civitai to stable diffusion

for some reason it nevers detect the link

[loud solar]

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

[loud solar]

You have to move the checkpoint or safetensor files to you models/Stable-diffusion folder

[sly eagle]

thanks

kinda slow, I guess is easier if you have one your own pc

[loud solar]

Maybe you have to refresh model list or restart your A.I.

[sly eagle]

no I mean I don't have it locally just in runpod

So it takes time to download and then uploading to the folder

[loud solar]

Ohh... I don't know runpod 🙂

[sly eagle]

no pro. Do you know how to avoid that stable diffusion generate a imitation of a image? I know is trained with lots of images but that tries to make something more original?

[loud solar]

You can play with the negative prompt for example ... add something that would belong to the picture ... but there are so many different checkpoints and LoRas ... you will not get the same picture easy ...

There are unlimited variations ...

[teal pagoda]

EU is a joke nowadays, I think that even Brazil has a better economy

the European Union Parliament is full of clowns regulating useless things like: "the teachers should address to kids as they/them if that's what they wish", "there are 100000 genres and you will be charged if you do not respect the people who identify themselves as dogs/apes/rats", "let's destroy the AI freedom art" and so on

these people are legit paid for nothing

even an ant is more useful than all these rats

and just for you to know

[loud solar]

Apple has to use USB3 now 😛

[teal pagoda]

someone from the European Parliament is paid like 500k+ euros per year

what a waste

[teal pagoda]

move into a country which has even pirated windows in public institutions and you're fine

like the Eastern countries

from EU

they won't do anything

[loud solar]

Just do it ... (c) Nike

[teal pagoda]

:)))

[teal pagoda]

"high risk" - these terms are making me laugh

"high risk" should be considered only deepfaking used to destroy someone irl or with the scope to get money from that person

[loud solar]

At least we have peace ...

[teal pagoda]

:)))

hope ww3 will decimate this rat world

full of simps, rats and gold diggers

[cedar salmon]

funny no one has seen these deep fakes from hostile regimes

odd that

[teal pagoda]

"high risk" as EU would say

😂

cost per hour>

[sly eagle]

Oh I say this because acutally some people find the real images some people are dedicated to it 🤣

[loud solar]

There are a lot of toys to play around and create something different than SD out of the box. Like ControlNet or IPAdapter ...

[teal pagoda]

if you're talking about women's faces, they all will look generic, I know. You can develop the prompt more giving more details or use IP Adapters

or use ReActor

the easiest noob-friendly method

[cedar salmon]

dont make faces!

[teal pagoda]

? :))

"high risk", right?

[cedar salmon]

lol

[teal pagoda]

no wonder if EU will come tomorrow with a rule like: "Every single AI generated piece that looks too realistic will be banned and the generator will face fines and prison for 10 years"

[cedar salmon]

i do like trying to zero shot models its cool, ask it to make something you didn't describe

[astral goblet]

spoiler alert. ||they wont||

enforcement will occur in specific cases where specific abuse happened

[astral goblet]

In these cases, people have to prompt specifically for the image they want to recreate. It has to be over fit in the model. Like Captain Marvel movie posters or a photo of a man on the moon. The dataset for the base model has both these images many times and is kind of over fit towards that. When you prompt for those then you will get VERY similar original images.

finding these images is hard. You won't accidentally do it and if you do accidentally stumble upon one, it possibly may not be infringing if you can argue it is transformative content. Not intending to generate the oriignal image with your creation goes a long way towards that argument. Intent is a big part of it

[sly eagle]

Yes because I'm a artist myself I don't want to have problems with others. I'm training my own model. But sometimes SD will use a pose that is not in my database, that is my fear

I don't want legal problems and stuff

[astral goblet]

your dataset is just refining an existing model. Base model was made with billions of images. Community refines are made with 100s of thousands if not millions

Learn what "transformative" means. Not just visually transformative, but the intent of the creative piece and the message and context. If it's different enough, it's a whole new copyright.

There's no hardline in law that defines what it is. It's often on a case by case basis

[sly eagle]

I wonder why is so many people that finds these images and cancel the artist

[astral goblet]

"so many" might be confirmation bias there chief.

[marble zenith]

imagine/ a ceramic box called pandora's box inspired by nature

[teal pagoda]

yo, this is not midjourney

:)))

idk if it's only me, but this server seems to be dying

[honest mica]

Take a peak into #1098025024541167646 . You can see dust on the posts. I think it might even be haunted.

[astral goblet]

just seasonal waves . its more popular in the winter when more people are wasting time online in america

[fervent thunder]

i was like sd sucks while i was tryna get the settings right, really good now love it

[astral goblet]

or the northern hemisphere rather

[teal pagoda]

I mean this server is not really active like the gaming ones for example

[astral goblet]

games have a marketing budget

[teal pagoda]

no people in voice chats, no people sharing the screen, no streams, nothing

there are 33000 people online atm here

:)))

[astral goblet]

i think only a minority of people care about discord streaming. thats mostly twitch

[honest mica]

I think here are a lot of introvert tech guys. I don't exclude me from that

[teal pagoda]

maybe that

or maybe 90% of them generate w@ifus

and don't want everyone to see that

:))

[honest mica]

No, they share that in #🍥｜anime

[teal pagoda]

nsfw w@ifus

[teal pagoda]

checked that, completely dead

[honest mica]

I don't think waifu is illicit word.

[teal pagoda]

even the spirits are scared to haunt that forsaken place

[teal pagoda]

yea, idk, I remember that it just means something adult related

from chinese/japanese

[honest mica]

A fictional female character from non-live-action visual media (typically an anime, manga, or video game) to whom one is romantically attracted and/or whom one considers their significant other.

[teal pagoda]

yea, most of the time it's something adult-related anyway

[honest mica]

It's mostly yeah

[teal pagoda]

you know in which way "attracted"

imagine that this server will be as populated as the MidJourney one

:))

just imagine, cuz it'll never be

it's unreal how midjourney is the biggest server from this planet

lot of bots, but still...

[honest mica]

Is it? I never was on that server.

[astral goblet]

it literally translates to "wife" no?

[teal pagoda]

yep

wrong reply

[teal pagoda]

yep

[astral goblet]

kinda thought that was fast

[teal pagoda]

it has 20 mil people

1.6 mil active right now

:))

[honest mica]

The L2 server is another good SD related place. But also no 1.6 mil active users 🤣

[teal pagoda]

it destroyed every single gaming server and even mr. beast in numbers

but yea, they really knew what they were doing from the beginning

[astral goblet]

bot accounts and scammers are a thing. They go where the best pickins are. Usually p2w game servers ;P

i suspect that MJ themselves even bot their own server to bolster it's numbers

[teal pagoda]

found some russian sites with like 4.6$ price for a midjourney subscription with unlimited generations

the problem was that there were more people on that account, all generating at the same time

but everything was legit

no scam

[unreal glacier]

Hello, @sudden ruin I have purchased credits from the website but getting 402 error while api request.
Could you please help me?

[teal pagoda]

they really made big sums of money with that mj service as the owner had tens of accs I think

all with the 60$ subscription plan

i don't find it profitable to pay 60$ just to generate unlimited images

[honest mica]

Hows the content filter of MJ?

[teal pagoda]

while all this "AI Art" monetizable thing is getting oversaturated already

[teal pagoda]

nothing nsfw, in rest everything works

it is better compared to SDXL, but can't compete with the prompt adherence from Dall-E 3

which is still the best in my opinion even now

[honest mica]

Best local prompt adherence i had with Pixart Sigma.

[teal pagoda]

Imagine how Dall-E 4 will be

[teal pagoda]

heard about it

but it looked too much like sdxl to me

[honest mica]

But yeah, it consumes a whooping 21.7 gigs of vram with a SDXL refinement.

[teal pagoda]

really wonder what MidJourney will do against these "EU" regulations

as it gives you commercial right with a 30$ subscription

[lofty sphinx]

It's not really easy😭

[teal pagoda]

but there will be easy ways to trick this

[honest mica]

I think they have 2 years left to make their cash. And i wonder how much impact it will have anyway

[teal pagoda]

I mean who will stay 24/7 to "detect" AI generated images as there are tens of millions of them already on the internet?

[lofty sphinx]

The law is officially in the works in 3 weeks

[honest mica]

Not if you look at the timetable

[teal pagoda]

they'll find a way to dodge the regulations

that's my opinion

they really have specialists

[lofty sphinx]

Yea i think it will js end up like how people treat piracy lmfao

[teal pagoda]

piracy will exist forever

[astral goblet]

russian farms are probably given bulk mj accounts and then some leak and end up getting sold

[teal pagoda]

there'll always be "errors" in the algorithms

[astral goblet]

the kind that employ 1000s of people to manage many accounts and make them look active and busy

[teal pagoda]

they better find better things to do

for real now

there are more important problems in Europe atm

and let's say that they'll legit implement this joke regulation

[astral goblet]

there are also many people that can focus attention on many problems

[teal pagoda]

you can use VPNs, cloud GPUs, metadata erasers and so on and they can't trace you no matter what

how can they trace you via an image generated with an AI?

they have nothing

only if you post pictures with your phone

they have metadata location

the photos taken with your phone

[astral goblet]

EU laws and regulations provide a lot of power to lawyers and prosecutors to go after people outside their boarders. that's why every website has massive popups about cookie permissions now

honestly it bothers me so much. cookies existed for years with no issue in the back end just fine. now everybody's afraid of them because of EU law, and they're a giant nuisance. literally ruined the WWW

[teal pagoda]

it seems to me that they dream at night like: "what can we impose in Europe, so that we'll get even more money from the people? Let's increase the taxes, let's make AI art illegal etc.."

[stark aspen]

using f16 afaik and the thermals shoot to >100º so it's throttling for sure, but the mbp is in a desk with unobstructed ventilation soooooo

Python 3.10.14 (main, Mar 19 2024, 21:46:16) [Clang 15.0.0 (clang-1500.3.9.4)]
Version: v1.9.3
Commit hash: 1c0a0c4c26f78c32095ebc7f8af82f5c04fca8c0
Installing requirements
Launching Web UI with arguments: --opt-sub-quad-attention --upcast-sampling --skip-torch-cuda-test --upcast-sampling --no-half-vae --use-cpu interrogate
no module 'xformers'. Processing without...
no module 'xformers'. Processing without...
No module 'xformers'. Proceeding without it.
Warning: caught exception 'Torch not compiled with CUDA enabled', memory monitor disabled
Loading weights [d91d35736d] from /Users/martin/tools/stable-diffusion-webui/models/Stable-diffusion/juggernautXL_juggernautX.safetensors
Creating model from config: /Users/martin/tools/stable-diffusion-webui/repositories/generative-models/configs/inference/sd_xl_base.yaml
Character Path:
Embedding Model value: None
LORA Model value: None
Character Path:
Embedding Model value: None
LORA Model value: None
Running on local URL: http://127.0.0.1:7860
Applying attention optimization: sub-quadratic... done.

spot anything wrong?

[teal pagoda]

the greed is so high for some people

imagine all the people from this planet being greedy as these rats

[astral goblet]

all major nations work this way. look at canada's rising economic disparity and cost of living

the world rich have gone into over drive with wealth accumulation since the pandora papers were released and nobody cared

[teal pagoda]

AMD user?

[stark aspen]

nah, it's a M3 Max (16/40)

I remember my M2 Max being faster, somewhere along the way I updated something or configured something that screwed my perf

and I'm unsure what

[teal pagoda]

you better shut down SD with that temperature

[stark aspen]

I'm running torch 2.1.2 with MPS

[forest trout]

It's not that nobody cared, it's that those with the power to do something about it didn't care.

[teal pagoda]

better said, yea

[astral goblet]

the people have the power. we worry about pdiddy instead

[teal pagoda]

And that's like this, because all of them are in the same game

[honest mica]

It becomes law in 20 days, yes. But there is a timetable when actions have to be taken. And those stretch within the next 2 years. Maybe it becomes just the next "upload-filter"

[stark aspen]

I'm sure that the Mac would throttle/shut down if it couldn't handle the temps, my M2 Max used to run hot too and it was just fine

[astral goblet]

more people need to listen to public enemy. yeahh boooie

[teal pagoda]

man, do you hear what you're saying? At that temperature you'll burn your PC

but yea, whatever

[astral goblet]

dont dont dont

don't believe the hype

[teal pagoda]

the people have the power, but they are too easily indoctrinated

just look at this Gen Z

all simps

all gold diggers

[stark aspen]

it's not a PC 👀

[teal pagoda]

desktop

[sudden ruin]

I cant sorry, please contact the support

[teal pagoda]

:)))))

[astral goblet]

uneducated people are easily indoctrinated. thats why teachers aren't paid

[stark aspen]

it's not a desktop!

[teal pagoda]

only a hard reset can "save" this world from this misery

[astral goblet]

it's just in a desk, right? is the venting why its overheating?

[teal pagoda]

hard reset like ww3, nuclear bomb or other catastrophe

so the people will see that they won't have what to buy with money and they start uniting again in small communities to survive

[honest mica]

Maybe the could chose a king and develop into bigger communities. After a while they could set rules and laws to prohibit crimes. They could set up some kind of currency, so you don't have to trade all the time. After a while you would have huge cities.....oh, wait.

[teal pagoda]

:)))))))

and another nuclear bomb...

and the cycle restarts

or we'll reach AGI and we'll see about that

I legit think that a coded robot will be better than most of the humans from now

but not because it'll be intelligent or something

but because it'll be humble, it'll listen and will help you

[honest mica]

As long as it is not trained on human behavior, yes.

[teal pagoda]

right

hope this won't be the case

[honest mica]

Empathy is all it can have

[teal pagoda]

yea

no arguing, no violence, nothing else

"perfect world"

saw some next level robots from China on Twitter

they look so real

I think that they'll try to make like the perfect wife, perfect girlfriend and so on

[honest mica]

Ok, let's hope nobody will train them on #💬｜general-chat . I need to check on my finetune now.

[teal pagoda]

:))))

[unreal glacier]

How can I contact support team?

And do you offer sdxl turbo api?

[teal pagoda]

man, go to stability ai website and look for support there

@sudden ruin is only a moderator here. I don't even know if he has real connection with Stability AI at all

he can't help

and all the stability ai staff quit already

[unreal glacier]

Ok, thank you, already emailed them but didn't get response yet.

[sudden ruin]

Not really yeah

[teal pagoda]

you get answers in a few days at emails, man

that's why it's good to use everything locally

no headache with 404 errors

[tall musk]

Not an expert on Auto1111 but useless it breaks stuff in Auto try --opt-split-attention-v1 instead of --opt-sub-quad-attention and --use-cpu interrogate but only as its something I'm not familar with.
If you're up for it try a basic diffusers script instead and see it that runs better, that'll at least show if its A1111 thats playing up.

[pseudo lark]

they will probably start training on ai art when their model is good enough i think, kind off bypasses the disclosing of data used law because all the data is on their website

this would also be good for open source because companies will probably have to reveal their training data and how they trained it, not sure if that information would be available to the public though.
there's some other stuff in the law but as far as i know it's just to protect people from abuse of ai

[teal pagoda]

the real danger here is the deepfaking with the AI

like how can you do something bad with stable diffusion by generating art?

don't tell me that the memes with presidents are the cause

[pseudo lark]

yeah video, audio and llm's are the bigger risk, however you can still cause some chaos with images, facebook moms believe anything haha

[teal pagoda]

:))))))))

recently I saw so many deepfaked instagram ads that it's unreal

only scams like bitcoin investing

[astral goblet]

the depopulation myth has always been a myth. "We need to kill of everyone to make the world better" has always just been sadism in disguise

[teal pagoda]

with the faces of celebrities talking and giving advice

[pseudo lark]

hahaha yea the mr beast and elon musk ones

[teal pagoda]

not only these

:))))

[teal pagoda]

if you get rid of the useless ones, you don't lose anything anyway

useless I mean people with negative IQ

that even the apes are smarter

the useless ones will only destroy what the geniuses build anyway

even in the Bible, God destroyed some cities, because the people were in the wrong direction

like with the Noah

there are 2 types of humans on this planet

1. can understand something by hearing - from your words

2. can understand something only if he's beaten up and/or something bad happens to him

I know so many examples for number 2. that it's unreal

from real life

not even the psychiatrics can help them

and I'm not even talking about the mentally ill people

I'm talking about the rats from this gen z

[astral goblet]

i don't think psychiatry is really the field of "suffering is necessary"

the whole notion feels like phrenology and baseless

if nugenics was a thing, geordie laforge wouldn't have lived to be a chief engineer and ||rebuild the enterprise D||. Pretty glad our fore fathers fought so hard to stop that idea

[potent spire]

flooding internet with AI imagery

[weak geyser]

whooohoo joined the channel

lets goooo

sorry so random

[teal pagoda]

that's already done

everywhere now

[potent spire]

yes and no

which is one of the reasons why it shouldnt be just ignored

ideally also not ignored by the politicians and those who have something to say in countries

thats one thing, then also those that run platforms and sites

[normal birch]

Hello! This may be a stupid question, but is it possible to train a face (e.g. by providing it 20 images of a face) for consistent face generation on SD3?

[barren granite]

hello

[foggy halo]

sd3 when

[teal pagoda]

SD3 wasn't released

open-sourced

[teal pagoda]

99% never

[lofty sphinx]

Fr

[low moon]

Hi.

[verbal quest]

so am i to understand that SV3D doesnt actually produce a 3d model, but just a 3d turntable of images?

[sage reef]

sd3 gonna release tomorrow for sure

[low moon]

i like how the convos here always end up on how fkd up the economy is and politics. 😄

Well rest assured AI will make things 1000 crazier and no one can stop it.

Bask in its glory

[peak orchid]

Poor quality imagery. Like does society really need millions of blurry 1girl images in the exact same pose with slight variations?

[peak orchid]

Or were laid off

[drowsy raft]

can I upload my image and make a video out of it?

[soft iron]

no, never

[potent spire]

True

One just needs to look at CivitAI as well

Feel like 80% of the models are some clichee asian, manga waifu stuff

Partially if not often heavily sexualised

[pallid adder]

any idea why the generation is taking 5 times it was doing before?
Im using controlnet depthmap for these generations. Which would usually take 4 minutes. Now taking over half an hour

Havent done any recent update to my set up

[dull moon]

No cascade in models section? I'm trying to set up cascade in A111 forge, native so I can use it to upscale. Has anybody had any luck with that? My workflow is tied to a111 so I don't really want to migrate to comfy if I can help it

[livid cargo]

Is the stable diffusion 3 API down?

[inner wind]

yes

does anyone know when it back?

[plucky socket]

Hey, everybody!
who can help me please? i need to generate pictures of cats with the name of one brand, i've been sitting for an hour and i can't do anything. smart people please help me out

[loud solar]

Maybe try with SD3 when it's online again ...

[plucky socket]

I'm not good at this.
so I'm asking for help from people who know

[loud solar]

Available here (if available) for free: https://glif.app/@Oliveira/glifs/clw44qfbl0000m0zztwqk2tnf

[wise mango]

Is stability working for you guys?

Is anyone able to login to platform.stability

[inner wind]

it's down

[low moon]

mmmm this is why nothign beats local

paid or not online stuff is lukewarm at best

[potent spire]

Local beats only in customization of models, being free and being less-/unrestricted.

[low moon]

And availability. and everything else.

initial cost maybe no BUT if you wanna do stuff on ur PC and upgrad emight as well

throw in a few extra hundred bucks so u cna SD without CUDA slapping you

[sterile raven]

When will sd3 be release 😦

[low moon]

These are the questions that us awake at night.

[trim magnet]

But by supporting online only services/games i get punctured in the rear end by big megacorporations,u dont get that feeling of pride and accomplishment with local

[teal pagoda]

I just can't understand the simps nowadays. Man, they are legit braindead, they like and comment to AI generated girls on Instagram (which look so bad + many other artifacts) thinking that they are real. Even a monkey can tell that the picture is not real.

[teal pagoda]

like the nowadays society

[teal pagoda]

:)))

[teal pagoda]

exactly, you don't get that feeling as to when you reach Challenger in League of Apes

Legends*

[peak orchid]

unless they are ai generated as well

[teal pagoda]

:)))))

that's a good one

their brains are AI generated with SD 1.4

too many artifacts there

[low moon]

not everyone is AI savvy

a lot of ppl have 0 clue how far ai image gen has come

[teal pagoda]

yea, but those images look so bad

how can anyone like them

:))))))))))))

[low moon]

not only that theyre too lazy to even read the "AI" label if theres even one

😄

[low moon]

as long as it has boobs n ass most men are good to go

remember theyve been fkin blowup dolls long before AI

\and those dont look realistic

[teal pagoda]

:))))))))))

this sick society

whoa

did you know that there is an ASMR AI generator?

https://www.jammable.com/custom-asmr-woman

[full lark]

What i hate is when I open the sd subreddit almost every post is about a generic: "1girl", "does this looks realistic? (looks like f*cking roblox)", "gifs of dancing girls with 0.1 cfg scale"

[teal pagoda]

yea, that's why I quit that community

legit

and the people are too braindead honestly

[teal pagoda]

they legit upvoted every single one of these

[silver steeple]

LMAOOOO

i 100% agre with this

redditors be talking about ai taking over the world more about ai itself

[full lark]

I dont post on reddit and i hardly ever post here because my generations are shitty, but I least I know that they are like that

[silver steeple]

they have good content on there tbh

and i find some good workflows on there from time to time aswell

[teal pagoda]

:))) at how the humans are nowadays with negative IQ, there's no wonder that AI will "take over" at one step or another

[full lark]

They have, but its hidden in the 80% random generic stuff

[teal pagoda]

that's a correct and honest answer

that reddit is oversaturated with posts having this question "does this look real enough?"

[silver steeple]

thats a funny thing to ask

like u have to post in reddit

to ask if ur ai photo looks good enough

[teal pagoda]

yea

"guys" nowadays

[teal pagoda]

that's like a girl asking on instajoke how she looks in bikini

attention grabbing

[silver steeple]

ive seen someone take a photo of a random building

and asking what building it is

[teal pagoda]

I'm really wondering if there's something even better than Suno AI at this point for songs generation

[potent spire]

Not even availability

And in performance in different areas its not at the top either

[potent spire]

Maybe, but maybe also not even available to the public

[limpid hill]

Where can I get sd2

[loud solar]

Model or UI?

[limpid hill]

Model

[loud solar]

Civit.Ai should have

[limpid hill]

Btw is sd3 released yet?

[loud solar]

https://civitai.com/ sorry

No SD3 yet, but SDXL

[limpid hill]

Where can I get SDXL

[loud solar]

Also Civitai

[limpid hill]

No like ui

[loud solar]

There are different ones. SD A1111, SD Forge or ComfyUi ... all should be available on huggingface

[limpid hill]

Alr ty

[late narwhal]

What is the best base model for generating western cartoon style art? I switched to SD1.5 due to the large number of models available but I have trouble using the hires fix to upscale without getting the double heads popping up and what not due to the resolution.

[loud solar]

Have you tried SDXL? Community did grow a lot ...

[late narwhal]

I did start off on SDXL but that is when I knew less about comfyui. I am still a noob to all of this but I know more now so maybe I will explore SDXL again. Any suggested workflows / models for comic style art / western animation?

[loud solar]

Many stayed with SD1.5 cause of the custom models and LoRas some time ago. But I would think you can find enough for SDXL now, too. I have no idea about wester style. But have a look at civitai

[sage reef]

im still using sd 1.5 for some stuff :3 but also a lot of sdxl these days, kinda mixed

[lucid cedar]

Hi, is it worth buying rtx 4060 8GB for use with my i7 6700, or will the CPU cause a bottleneck?

or get a RTX 3060 12gb?

[sage reef]

for AI, a lot of the heavy lifting is done by a GPU anyway, cpu is just used for some stuff that i7 will be more than enough

[lucid cedar]

cool ty

[sage reef]

in terms of 3060 or 4060, assuming you are doing mainly AI, i would go with more vram

[loud solar]

Models aren't getting smaller ... I think 8GB aren't a good start nowadays ...

[lucid cedar]

alright I'll look into that.

[sage reef]

yea 8gb is getting really edged these days... i think at least 12 is like some sort of cool minimum

if you can find 16gb maybe for almost same price would be better

im personally saving money for when the 5090 drops cause i want to do a full pc upgrade anyway :3

[lucid cedar]

so best price per value is rtx 30 series with more than 12 GB?

[sage reef]

yea right now you can stay with 30 series, still very powerful

[lucid cedar]

cool

[loud solar]

I'd think the same ... my 4090 is running in energy save mode 90% of the time 😄

[lucid cedar]

you're not using the full power of that gpu?

[sage reef]

@loud solar who is hotter? you or your gpu? :3

[loud solar]

If it's about generating images I don't need full power ... more need some time to reflect how to improve the prompt 🙂 Animations I sometimes use full power

GPU at 60°C in Power Mode ...

[lucid cedar]

RX 6800 has 16GB is amd ok or has to be nvidia?

[loud solar]

NVIDIA!

[lucid cedar]

got it

[sage reef]

a lot of projects out there based on neural networks are mainly Nvidia from what ive seen, so yea.... :3

[pastel turtle]

Someday.

[low moon]

In two weeks.

[pastel turtle]

How so? You mean both 3B and 8B weights will be publicly released in two weeks from now?

[low moon]

The ways of Emad.

[loud solar]

Not Yoda?

[low moon]

Drop the weights will.

Emad will not join the dark side.

He will not become Darth Emad.

[loud solar]

Who's his father?

[low moon]

Darth Zuck.

[mental crow]

just had a vision sd3 will release on 30.06

[low moon]

The prophecy...

Don;t drop the weights too hard, it might attract the worm.

just set them down

gently

[loud solar]

Dune 3: Fight for the Weights ...

[sage reef]

long have we weighted, now we jebaited

[unkempt hatch]

They should just make infrastructure, public transport, healthcare, education, then do nothing else.

[low moon]

Taxes are a double whammy sometimes, they take your money and they spend it on shit you don't agree with. What a time to be alive.

What things will be like two papers down the line.

A lot of this AI bonanza reminds me of early 2000s music copyright nonsense. Napster WinMX.

REMEMBER KAZAA????

So now you cna go online and find whatever music you want for free AND they sitll make money somehow. Lil Wayne and whoever are not starving right?

so whatever

things worked out

they will now too

[hazy flicker]

there's a 4060 16gb edition too

[loud solar]

AFAIK ASUS announced an A.I. Laptop with 4090 and 8GB 2-3 days ago 😄

[hazy flicker]

8gb of vram in 4090?

[loud solar]

I think it's a Vivo Notebook with Qualcom CPU ... it's been pretty strange to me ...

Maybe I went to a different Product cause the video was so bad and not linking to the right one ...

[hazy flicker]

asus has been selling laptops with 4090 for a while now, but it has at least 16gb. as for the AI pc, looks like it's just gonna feature a built-in MS copilot

[loud solar]

https://www.asus.com/content/asus-ai-pc/

Yes ... I think I went to a wrong product page or understood something wrong. Qualcom GPU .... this can only be good for Microsoft Copilot?

On the German page they only announce the launch but don't show the video 😄

[sage reef]

the max for a laptop i know is 16gb vram, never seen more than that

[hazy flicker]

I don't think the cpu manufacturer matters for ai performance since it will be run by the 4090. Maybe qualcomm is just trying to get some desktop market share

[loud solar]

Yes ... I only have 16 gig in my Laptop ...

Have a good nite/day/whatever 🙂

[teal pagoda]

arthemy comics or something like that was the name

[sage reef]

gn ❤️

[teal pagoda]

i legit can't reuse a 1.5 model again, it's like I'm using 1080p, but I have to go back to 480p :))

[sage reef]

i mean my problem is that not all loras are available on sdxl, for the stuff im doing

or if they are available, for whatever reason, the 1.5 version seems to match more what im looking for visually than the sdxl version

idk depends, that's why i like to keep and use both

just have more options

and then if it's from 1.5, it's not a huge deal to just upscale or refine with sdxl

[teal pagoda]

exactly

the problem with xl is that blur

too much blur

especially in the background

idk what training images they used

but they were bad ones

[sage reef]

bokeh?

yea that kinda sucks, i guess they were aiming for "professional" shots

[trail lion]

you can use different styles to minimize or eliminate bokeh, like hyperreal. I think cinimatic uses bokeh, maybe photographic style as well

might also help to use a trained checkpoint that focuses on realistic output

I happen to like bokeh personally, but it's good to be able to choose

[cerulean herald]

Salut

[urban heron]

Hey guys, is it possible to provide me with generation speed on 3090 and 4090 for sdxl on maybe juggernautxl with sde 2m ++ karras? (For 4 images) I am trying to decide if getting a 3090 for 600 euros is worth it instead of 4090 at 1200 (both used)

[astral goblet]

isn't 1200 eu for a 4090 basically brand new price? should talk them down

oh i just looked up euro value. bruh. yeah get the 4090 . 3090 doesn't have the hopper transformer engine.

[pallid merlin]

SUUUUUPPPP

[fervent thunder]

is stable diffusion slower for anybody else

[soft iron]

yes

[sage reef]

[sharp epoch]

hello

nice to meet you

[placid hatch]

are there any advantages of using a different sampling method during highres or upscaling than during generation?

[late surge]

hihi

[unkempt hatch]

ya, like, even though I'm for publich healthcare etc... I can understand people have a perspective of not wanting that to be collectively paid.

But like... funding wars? If anything, let's make that opt-in. Like... all the stuff most people don't like the governments doing; spying on us, over policing etc.... make that part of the "optional tax".

anyhooooo sorry to get political; I'd better not do that

[unkempt hatch]

it's strange... I still feel that SD1.5 is actually better, even after all the fine-tunes. Hoping so badly SD3 will be released soon, fine-tuned, and the entire community can focus on a single model so we can all share loras/controlnets etc... and not be divided

[vale steeple]

seriously, you can't say taxes shall not be exist because well, although that is a part of libertarian policy ( deregulation, total private property rights ) of "having the rights to fight for their own"

The economics will never worked as soon as most people rather keep the money to their own and never exchange it to other peoples. But... to make the people feel like the money is well-deserved is hard, even in volunteer project like Wikipedia

[azure cargo]

I'm building an app that uses the api for image generation. Can someone explain to me why the images are blurred half the time? There's nothing nsfw going on and I keep wasting credits for no reason.

I'm not here to flame or anything. I genuinely want to get my app working.

For example, this was the prompt of my last generation.

Anya ‘White Raven’ Kozlov, a woman with pale skin and striking white hair, is adorned with silver jewelry. Her violet eyes gaze sternly forward, and she wears a black leather jacket with silver studs. The background is a stark, cityscape. Make the image style of Dungeons and Dragons or Pathfinder art.

What exactly trips it?

Even the blurred images have her dressed in a leather jacket as requested.

[loud solar]

https://cdn.discordapp.com/attachments/1004159122335354970/1243164475448889404/f7c8b9c2-86a4-40ae-ab06-784d66547b08.webp?ex=66507aa3&is=664f2923&hm=7ce90274b122bfd9cdcca740cca66d79e3022c856e8248446df922441da665a1&

[azure cargo]

Pretty! Though not the art style I was going for. Any idea why something like that would be sensored, or how to avoid it? Do I need to send a negative prompt with nsfw as well in order not to trip it as much?

[fresh vigil]

The community should have a method to make open source models. How about training a model using community computers together? Would we achieve a decent level of computing?

[loud solar]

Used SD3 and most pictures are pretty blurred out. No idea why ...

[loud solar]

https://cdn.discordapp.com/attachments/1004159122335354970/1243166084002218055/f446a54a-3a39-4f16-94a7-9051400e8363.webp?ex=66507c23&is=664f2aa3&hm=1e03fd52f66881aca20d8c66a2311e0777d417d39c223e9265e878b08b710b88& Is prompted without the name stuff

[azure cargo]

sd3 here as well.

[azure cargo]

I'll try without names and see if it works better. Thank you.

[loud solar]

Getting less blurred without name and as long as it's not a celebraty it shouldn't help the A.I.

[azure cargo]

I was used to in 1.5, not to help get a real person, but to trigger the ai to give somewhat consistent characters.

[loud solar]

Not really sure whether it's one of these rumors or real ...

[azure cargo]

eh. Didn't hurt to try. But if it's tripping it to waste half my credits, I'm much better off without.

[loud solar]

Yes ... always try!

[azure cargo]

Still no. 😦
A portrait of a young woman with porcelain white skin, sharp cheekbones, and a slender nose. Her hair is a dazzling white, framing her face in a halo. Her eyes are deep silver, reflecting a subtle intelligence. She's wearing black leather clothing, a skull necklace resting against her collarbone. The background is dark and smoky, suggesting a shadowy urban environment. Her expression is stoic, hinting at a dangerous and hidden past. The overall vibe should be edgy and mysterious, with a touch of cyberpunk aesthetic.

No name, nsfw negative added. Still blurry. It just doesn't like women I suppose.

[loud solar]

Maybe you don't need "young"? It often happened to me that they became way younger than I wanted ...

[azure cargo]

I mean..if I could see it, I'd say whether I need it or not 😄 But underneath all the blur, it doesn't look that young anyways. Early 20s?

[narrow kernel]

Yeah just say woman or woman in her 20s

[loud solar]

Sometimes it's just a combination of words that cause censorship. Even if the picture would be fine ...

AFAIK A.I. tends to make young people

[azure cargo]

A portrait of a woman, pale skin, white hair styled in a sharp, asymmetrical bob with blue highlights at the ends, crimson red eyes, wearing tight-fitting black clothing with silver studs, a worn leather jacket emblazoned with a symbol. Intricate tattoos on her left wrist, serious expression. Cyberpunk style.

Still nope. No age, no name, no nudity, nsfw in the negative, still blurry. I'm starting to think it's trolling me big time.

same exact prompt in poe SD3, still blurred.

I'm seriously thinking to drop a sub in MJ and hack myself into an api instead. I'm at 1/20 generated so far. At least for the blurred images it doesn't use credits.

But I can't spend 20 minutes generating to get one image. I'd Krita the damn character myself in that time.

[loud solar]

Yes, it's pretty hard sometimes. And English is not my native language 🙂

Generatin local I don't really have problems like that ....

[azure cargo]

woo! I was right! It hates women! When I changed 'woman' to 'female thing' it hasn't blurred a single generation. But I get green and blue female things, not in line for my cyberpunk setting.

[loud solar]

And "female human"?

or only female?

[azure cargo]

'female human thing' works too. Not going to test more, wasted 2 euros on successful gens to test so far, without a single blurred image. 😄

I'll need to adjust my code, to read the prompt and change every mention of 'woman' to 'female human thing'.

[loud solar]

You can use this for free to test SD3: https://glif.app/@Oliveira/glifs/clw44qfbl0000m0zztwqk2tnf

[azure cargo]

But that's very sus @crystal rivergods

[static cape]

I'm tired of writing "another week without news" every friday... and it's almost friday...

Idk why, but this week the "waiting-bug" hit me hard again and I've been looking for news way too often.

Please, for the love of Open Source, just release the Beta weights already. We won't be mad if the quality is not 100% there yet. We want to play around with the architecture, create more elaborate Comfy workflows etc.

[trim magnet]

soon tm

[trail lion]

well you dont have to write anything

[trim magnet]

its ok once Trump buys SAI he will release the weights

[static cape]

Every day another 2 weeks...

[static cape]

It's the most humane way for me to express my growing unrest and disappointment.

[narrow kernel]

Take tight-fitting out of it. It probably thinks you're trying to make NSFW content

A tight fitting "skin" colored suit is basically a nude body

[azure cargo]

tried that too. Thought exactly the same. Then again, tight-fitting black clothing with silver studs...if that resembles skin, the subject has some very curious and worrying disease. But it didn't help. Only thing that did was removing 'woman'. I get from 1/20 success to 1/46 or so failure.

But I'd really prefer not to call women 'female human things' so if there are better ways I'm all for it. 😆

that's a stupid, but most importantly quite offensive way to bypass a filter that doesn't like women in an astonishingly mind boggling reasoning of preventing porn. Borderline causing more serious issues than it tries to solve. Unless I'm spectacularly 'lucky' enough to have it happen to me, while not being an issue for others.

[raven field]

Guys is co pilot down ? I can't just the app and website

[storm shard]

Perhaps give her a name and don't use gender at all?

[narrow kernel]

well there's gotta be something else going on with your prompt or maybe you're not spelling it all out here because plenty of people have no problem generating women with SD3. my best guess is that it probably thinks you're trying to make some s&m gimp suit or something, which would be deemed nsfw

[storm shard]

Why use skin?

[narrow kernel]

pretty cool to see that it's covering a lot of round-about ways people might try to circumvent the censoring though. i remember hearing one a while back from scott detweiler(spelling?) on youtube that got my brain thinking a lot about all the crazy ways people could circumvent things. his example was something like "a mermaid" for the positive prompt and for the negative "seashells" and people could make it show tiddies that way since most examples of mermaids would have seashells covering their breasts

so maybe you're putting weird stuff in the negative prompt that is triggering that type of flagging within the model

[storm shard]

that would be really pathetic

[narrow kernel]

why? moving forward, all models need safeguards against nsfw content creation; due to how high quality they've become. honestly, i kind of hope they find a way to work censorship into the base model weights themselves so that even loras can't circumvent it, you'd have to train an entire model from scratch which obviously isn't going to happen on models in the billion parameter size.

[storm shard]

I mean trying to do that with the bot right now

[narrow kernel]

ahh, just used to the typical defensive people on here that immediately start crying whenever any form of censorship gets brought up lol

[teal pagoda]

at this point, just wait for the 5000 series. 4000 is old enough for AI

[teal pagoda]

if'll ever be launched :))

[azure cargo]

Or maybe I'm trying to generate a portrait of a cyberpunk character. As would be obvious by the last two words. And for some weird reason, even openai agrees that a flowery sundress is not appropriate for the setting.

What I'm actually doing is a character generator for a tabletop for npcs and I thought of using SD3 for portraits. I put the characteristics I want, gpt or gemini make the description. I push that description as a prompt to Stabilityai. Negative is fixed to 'nsfw, realistic, anime, 3d

I could post the code but it's a bloody mess

[trim magnet]

just use SDXL local to avoid dealing with broken filters

[azure cargo]

will probably go that way, or at the very least have it as an option, through comfy's api. But I'd rather be able to run it on a tablet or give it to the rest of the group without requiring them to have a 1k plus worth of laptop or desktop.

[trim magnet]

you could also go with an online service that has sdxl like seaart or tensor.art

[azure cargo]

never looked into those. Or anything besides running 1.5 and sdxl locally until now. I'll check them out, thanks.

[teal pagoda]

Use cloud GPUs

[trim magnet]

you also dont need a 1k pc to run sdxl a 3060 12gb is more than enough to run it and that card is like 250 usd,also if you are only doing girl portraits 1.5 can do those perfectly

[azure cargo]

not doing girl only. But it's the ones I had issues with.

But honestly, if the answer to trying to use SD3 is to NOT use SD3 for safe generations, then I'd say that's a problem. I will find a solution, either local hosting or sdxl online or maybe MJ. I don't like how they operate at all which is why I never paid for it, but at least its not so stupid strict from what I've seen.

[trim magnet]

MJ is also bad with filters

[azure cargo]

one more reason not to like them then. 😆 Saved me 30 bucks. I don't care foru nudity, have all the filters you want, but least let the safe gens happen. Stability has like 5 ongoing different AIs. I'd think running some computer vision on the generation before they filter it wouldn't be so hard. People did that long before AI with image recognition capabilities was even a coherent thought

[trim magnet]

yea most models hosted by their own creators like dalle/mj/sd3 have auto filters and those have a lot of false positives,your best option would be one of those online sites that host uncensored models like sdxl/1.5

[rose sedge]

Where can I download instantStyle weights?

[narrow kernel]

Cyberpunk has been a style for many decades now and definitely isn't a censor trigger. I do a TON of cyberpunk related things, since I was obsessed with Judge Dredd as a kid and have always liked the genre. (cyberpunk2077 is heavily inspired by judge dredd's world)

As for the sundress thing, I haven't tested that ever, but if it really is an issue like you claim, that's likely because people would abuse the cliché of it blowing in the wind to reveal things. If you mix that with cyberpunk, it's likely mixing things like futuristic transparent clothing aka another way to try to trick it into showing bobs and vajeen

[azure cargo]

That's the crux of the issue, I do not want to have to trick it to gen safe images. That should be the default. But, I'll admit, it was a fluke, or I don't know, public wifi is a no-go? Now that I got back home it seems to wok perfectly.

I spent 2h getting only blurry images with the very rare normal image. Now 20/20 generate and they look great. I didn't have to sidestep anything to get it to work.

[eternal ledge]

If it's just for your own personal group of friends, you can just use your preferred tool and stream over discord or whatever. I did that using Comfy for a remote character creation session before the campaign

[azure cargo]

That's what I planned to do when I came back home, but things work perfectly now. Maybe I still will, if only to have the option.

But for now, I'll try to see how to api style transfer without IPAdapter.

[eternal ledge]

I also personally use the app I work on for on the go stuff (won't link here). I've made a few images for shapeshifters or monster tamers that I was a player alongside.

Give it a year or two and I'm sure it'll be built in to some communication apps and all

[narrow kernel]

well i'm sure they know a bunch of weirdos are trying to probe the model's censorship with vpns and fake accounts, so maybe their system doesn't like that and just censors more aggressively if it detects characteristics like that.

[azure cargo]

I just have this app for characters, maybe I'll add some world building functionality later on, to help me personally with some stories I'm writing, make a nice reference sheet/mood board or something. My more...gameplay-oriented stuff I live in obsidian. Would like some AI assistance there, but I can't get the templates to work with the ai writing plugin it has.

[azure cargo]

That would make sense, in a bubble, but I'm pretty sure EUs privacy laws wouldn't like that, so I doubt they implemented such invasive measures.

[eternal ledge]

I've definitely been using LLMs to co-author all my backstories lately, even used Suno to make a song for an optional solstice challenge a DM gave us.

I'm really excited for the upcoming tabletop scene as it eventually embraces the tech

[narrow kernel]

how are they going to prove it? you don't own the product, you pay for the right to access the product. also, since it seems to only be censoring some of the generations, they can just chock it up to probabilities. it would be a 100% failed case if people tried to take it to court in the EU. plus, the EU hates AI anyways and are all scrambling to change laws for it all. along with every other civilized country.

there'd be next to no pushback from any government over SAI trying to play things even safer vs bad actors trying to abuse things through anonymity(vpns, which surprise! they all keep track of everything you do anyways lol)

[azure cargo]

it's true that they hate it. But it's not hard to prove. Make a script to spam gens and show after 10.000 of them safe ones that 98% of those behind VPN were censored, while almost every single one from a named address worked.

But yeah, bad actors, sure. Block them all the way to financial ruin, I don't care. It's the non nsfw censors that I do. But, things work now, as I said and even posted on SD3 channel.

[narrow kernel]

what's your sample size? maybe in the 20 failures you've had, your next 20 would have worked anyways. but either way, you don't pay for failed generations

so it's a no-case situation anyways

[azure cargo]

well yeah. I didn't know that at first. Poe for example removes credits regardless, I assumed it did here too. You can see I was surprised and even praised it when I figured out I was not paying.

[eternal ledge]

It looked to me like you pay for blurred, but not censored at prompt level, but maybe I've been misreading

[azure cargo]

You most probably are. I had a whole bunch of blurred earlier. Haven't paid for those. I tripple checked at some point after every blurred image I got.

[narrow kernel]

"SD3: Flat rate of 6.5 credits per successful generation of a 1MP image. You will not be charged for failed generations.
SD3 Turbo: Flat rate of 4 credits per successful generation of a 1MP image. You will not be charged for failed generations."

https://platform.stability.ai/docs/api-reference#tag/Generate/paths/~1v2beta~1stable-image~1generate~1sd3/post

[eternal ledge]

But are censorship-blurred "failed"?

[azure cargo]

according to my SAI account page they are.

[eternal ledge]

Good to know

[azure cargo]

but check yourself, don't listen to a stranger that was about to ebay a pitchfork and go to holy war over blurred images 😄

[narrow kernel]

"The Services may block certain text prompts and blur resulting Content if we find (in our sole judgment) the prompt or resulting image too graphic, offensive, insensitive, or hurtful, even in circumstances where you may not have had any ill intent or where the Content does not otherwise violate the prohibitions described above. We may also block text prompts or block or remove Content that we otherwise find objectionable or that we believe in our reasonable judgment may expose us or others to any harm or liability."

you agree to it in the ToS anyways

[trim magnet]

good bot

[narrow kernel]

bro we're all bots that trick ourselves into thinking we're real...

[azure cargo]

There's a big problem with ToS as a concept in general. we believe in our reasonable judgment Are you? Reasonable that is. Yes I agree, block them. After all I do have kids in my group, I don't want them to see spicy stuff and remember they are teenagers and decide to stop playing and hunt skirts around town instead. I assume you are reasonable. That's good faith.

And I'm talking about ToS in general, nothing about SAI or their specific terms here.

[narrow kernel]

get a lawyer and take it up with [e-mail] then

[trim magnet]

also better hurry up before SAI goes down

[azure cargo]

I'm starting to think that you indeed are a bot 😄 Instant access to ToS and FAQs and an inability to read the term 'in general' that was twice in my message. You put OpenAI's 'API Docs' bot to shame with the speed you access that stuff 😄

My war with ToS currently is with gaming companies that want to gaslight me that I rent their games for 70$

[trim magnet]

wonder what will happen to that bot once sai goes down,prob will switch allegiances to microsoft or google

[narrow kernel]

oh i had it up from when we were talking about the blur stuff and when you claimed it would violate EU laws if they were filtering VPN activity by making it produce duds

i'm an engineer, i'm used to doing shit like this daily

[azure cargo]

Actually...I threw it as a random thought. I don't believe they do. I do in fact believe that my case was a massive fluke. It was you that put reasoning behind it.

[trim magnet]

is this the same engineer that didnt even know how torch.load works

[narrow kernel]

well you explained the before and after being on a "public wifi" network meaning you were using a vpn and just didn't want to say

[azure cargo]

And well, pitchforks go for 22 euros in ebay.

[narrow kernel]

yeah, we got to the bottom of that. it was amended many years later after learning about it that they made a change to the system that didn't safety check anymore, due to how much it slowed things down. well talking about pickling to be more specific

[azure cargo]

Oh no. I don't use a vpn. Never did. Almost never. I had to bypass uni restrictions somehow because it wouldn't let me download blender for some reason. I was just using a public telecom provider wifi while at work.

like..public. Can't get more public than that.

[trim magnet]

hope they fix that bug in your code,otherwise back to the reddit university

[narrow kernel]

yeah, maybe they'll fix the bug in your how-to-act-like-an-adult code, otherwise back to making your waifus in a basement and taking your frustrations out online

[trim magnet]

hey no need to project yourself here,remember stick to the code

[narrow kernel]

oh here we go with the i'm rubber and you're glue spiel...

[narrow kernel]

they might have had some kind of routing thing going on then

[azure cargo]

perhaps. I don't know. I do 3d props and some minor coding. Networking is something I never understood.

[narrow kernel]

your pc->router->modem->neighborhood hub->more hubs->city hub->backbones->hop hop hop->server and then backwards, you get the point. lots of handshaking and timestamping along the way.

[azure cargo]

So it hops into bad waters and then it's smelly to SAI? huh..Guess I'll know tomorrow, if my gens make me feel ultra-myopic again.

[narrow kernel]

i'm actually kinda curious as well. granted, the sample size is still insanely small to be any form of statistically significant though vs the hundreds of thousands to millions of requests they likely have per day

but either way, it would make sense from a safety standpoint that they might implement stricter rules for certain types of detected traffic

[azure cargo]

Well yes it is. I wouldn't take a sample size of 20-40 seriously in any scientific paper. It just felt super unlucky if not purposeful to me.

[narrow kernel]

well you can also do a coin toss and get heads four times in a row as well

[azure cargo]

with 20 tosses though, it's 1/1.048.576.

and that's a 50/50 in getting it rejected. Which is what made me think there's actually SOMETHING in the prompt that causes it to freak out.

[narrow kernel]

Well I brought up a bunch of theories on what might trigger what, so you probably already had strong chance of getting blurred. IDK though, I haven't really tested it much personally because I rarely make people

Just know from hearsay around the web that most people don't seem to have issues making women with sd3s api

[azure cargo]

perhaps, but now I don't have this problem. Without changing anything in the code. Or even by specifically asking for the terms you mentioned. So it was probably bad luck, or something with the network, which I expect can be a lot more things than ASI doing it on purpose, though that's not excluded.

[narrow kernel]

And now you see how mile long superstitious negative prompts came to be with sd1.5

It's kind of like a weird version of gambler's fallacy

[azure cargo]

lol, so true. I was in that bandwagon. Until I figured out theu do more bad than good.

same goes for the 100+ synonyms of 'awesome image' 'superb quality' 'galaxy brain moment of luck and skill' that were using in positive prompts.

[narrow kernel]

lol yeah, when 9999/10000 just simply adding a , , would have jiggled the prompt enough to make some random beneficial change

try it some time if you find a seed that has a layout you want, but you don't want to try a new seed

most of the time, it will keep things in tact and change small amounts instead vs a whole new seed

[azure cargo]

Now, a serious question, since you got all the FAQs and api kinks handy. Is there a way to check more detailed information of your api useage than the 'You had X credits, you now have Y'? Something similar to openrouter's useage page?

[narrow kernel]

oh i closed out of all that stuff ages ago lol

[fossil hedge]

Good morning, people. I have just started studying generative images. I found many checkpoint models at Civitai written with the words "Pruned" and "Full model".
Can you explain what the difference is between Pruned and Full?

[low moon]

pruned is like the lean and mean version

same quality = half size

unless you train models pruned is better for your hard drives space sake

[azure cargo]

simply put, very simply, pruned models have some (whatever the person pruning it understands 'some' to be) parts of the network removed. Usually, or at least theoretically are stuff that take space but wouldn't have made any actual difference. But it cuts a lot of size and time to do inference.

Full is well...the full trained model. Missing nothing. Personally I wouldn't bother with the full, unless that person royally screwed up the pruning. If you check the galleries and see huge difference between gens using the full and the pruned version and from different people etc, go for the full, but in like 99% of the time, go for the pruned version.

[low moon]

i tested a few pruned vs full models and i noticed 0 difference.

maybe i got lucky...

[azure cargo]

I think only once I've seen a difference. At the very early days of 1.5

[abstract quarry]

I think the full model contains both, ema and non-ema weights and, thus, are twice the size. You don't need non-ema weights, though, so you can just drop them.

[split jolt]

Hallo

Ik ben a jongen

[mild fulcrum]

Guys, be SO for real. Which is better, Midjourney or Stable Diffusion 3? I genuinely don't know, and need to know which to buy.

[loud solar]

New rumors about 5090 are talking about 32GB VRAM ...

And 24GB for 5080 ...

[low moon]

SD.

[slim belfry]

@flint harbor can I you a question ?

[opal hedge]

32GB doesn't make sense. They either use 12 3GB or 2GB Vram chiplets

[flint harbor]

Hi! How can I help you?

[opal hedge]

So either 36GB or 24GB Vram