💬｜general-chat

[warm junco]

ohh okay

[sick venture]

Any tips for ComfyUI?

[warm junco]

maybe use --normalvram

[woeful island]

Hello @everyone! I Hope you are doing great!

I am getting into stable diffusion, was thinking if there is a well structured course for it available! can anyone help me guide to it!!

[fervent thunder]

Could anyone give advice? I have a gtx 3060 with 12 gb vram. I’m trying to train embeddings in SD but it tells me I need more vram, even though it doesn’t seem to actually need more than maybe 8gb for the training.

[remote cosmos]

hey guys, im not sure exactly where to ask this but when you want to train a LORA on a particular art style, do you just give it a bunch of images in that art style and tag as normal?

[vague pond]

well, i've discovered a horrifying problem with sd1.5 and its VAE

LDM didn't train it correctly, and as a result it's smuggling information about the whole image in a few specific pixels. you can tell by the fact that every image seems to have a hole where logvar is extremely high compared to everything else, and that the KL divergence is extremely high on the diagonal gaussian

i also have evidence that the unet has also learned this bad habit because latents generated by SD1.5 have some similarly fragile patches. looking at the mean change in the decoded image after setting a single pixel to zero, i can see that a few pixels cause massive changes across the whole image

SDXL's vae is trained properly and doesn't seem to have this issue, but SD1.x and SD2.x are probably being held back a lot by the parameter space that's being used to pack information into these few latent pixels. and isn't DALL-E 3 also trained on the same VAE encoder?

[vague pond]

We use the same VAE[9 ] developed by Rombach et al. (2022) for our model.
dear lord, they did use the same VAE. so DALL-E 3 almost certainly has this same issue lol.

[rich kestrel]

wut

1.5's been around for eons

and now all of a sudden it sucks?

[pine path]

vae trained wrong make worse model

[rich kestrel]

then use a different McVae?

[forest trout]

We are building off of the old vaes so it makes sense that we have made a lot of progress since then.

[pine path]

not that simple. You'd need to use a model architecture that's aligned to a proper vae or align it yourself which is very costly.

[forest trout]

That's funny that openai didn't train their own though.

[rich kestrel]

no need to fix something that already works.

[pine path]

it works, but not as well as it could

[rich kestrel]

well, we're close to XL being on par with 1.5

I wouldn't say we're there yet tho, but closer

[vague pond]

"Progress" isn't the word I'd use here... like, what CompVis did is objectively wrong, and this is something that should have been easily caught during validation. If you look at the log variance of the VAE's diagonal gaussian distribution, there is a huge, obvious, unmistakable hole in it, and the KL divergence of the latent is extremely high. Like... the whole point of the KL divergence term in the loss function is to prevent exactly this scenario from happening where the VAE tries to smuggle information about the whole image through certain pixels.

[vague pond]

As far as 1.5 goes though, it still very obviously works fine. It's just that the VAE is making its job unnecessarily hard, essentially, and because of that it could be better with a better VAE.

it's possible that the vae could be fixed and used on existing models with less than an entire from-scratch retrain, since the smuggled information isn't really that much

[forest trout]

Not sure if you remember that paper discussing a single pixel attack but I wonder if they were testing off of that vae.

[vague pond]

#🏞｜general-with-images message here's an example of the effect i'm seeing

[forest trout]

Fascinating.

[vague pond]

but just... how many models used KL-F8

sd1.x, 2.x, dall-e 3, SVD, what else am i missing?

[forest trout]

Midjourney partially used stable diffusions training at one point. Maybe they used it too?

[vague pond]

and i discovered this completely by accident while inspecting adversarial noise images

[forest trout]

Uh huh. Interested in nightshade are you?

[vague pond]

"hey, why is the kl divergence on this latent dist so high"

[vague pond]

i do believe in verifying their claims at least

[snow spear]

hello

I just installed automatic 1111

I am searching for a style file

to make creating easier

[hoary sapphire]

Is there any ways to disable NSFW Filters on SD?

[fallow veldt]

hi, I wanted to ask, what's a good platform for uploading some kind of portfolio showing different worklogs with little explanations. AKA as a webpage, but I don't want to make a webpage or get a server, etc. Is there a good solid options so I just can link it when searching for work?

[sage sky]

anyone know how to put stable diffusion on my pin tab

on bottom desktop

?

[terse anvil]

if its the start.bat file then probably make a shortcut and put it in some windows folder where all the start files are, i am not sure where, u gotta google it.
from there u can probably pin it when it is visible in the windows menu

[terse anvil]

github?

[terse anvil]

get models that are not filtered? or what filters do you mean?

[fallow veldt]

Can it be used for that, like a portfolio?

[terse anvil]

i dont know what you are really trying, but if it is for code you can show it there, changes can also be seen, maybe not ideal for workflow only, so it depends what you try to do

[opal hedge]

Hello, can anyone help me with regional prompting with sdxl in comfyui?

I'm looking for any sort of reliable solution

[fallow veldt]

No, no for code, I mean, some page where I can post images, a little explanation of posibles techniques, with reference image and generated images

[terse anvil]

maybe something easy like instagram? otherwise im thinking of some kind of blog website, there are probably some free ones

[fallow veldt]

yes I was going to say maybe instagram, and made some image like a webpage instead of full images

yes maybe that's the most straighforward, thanks

[terse anvil]

i made a start with it, but i never went through with it, how far did you get with it and where are you stuck?

looks like inspire pack has regional prompting, idk if it works for sdxl and for your use case: https://github.com/ltdrdata/ComfyUI-Inspire-Pack

also have you tried inpainting?

[wary maple]

Does anyone know of a community/discord where I can find entrepreneurs using Midjourney/AI art to replace product photography?

I'm looking to market their services purely on commission, but I'm struggling to find someone...

[terse anvil]

midjourney discord maybe, social media

[fallow veldt]

really afaik and test it regional prompter is complete garbage for getting multiple characters. Works decently for composition or prompting different parts of same character. Multiple characters is pure frustration. Inpainting the other character seems to be the only sane way, or just prompting multiple characters without it.

[wary maple]

I just started posting in the midjourney server, but nothing of interest has popped up yet

[opal hedge]

Huh, I didn't know inspire pack had regional prompting capabilities. Thanks for letting me know

As for inpainting, it's a lot of busy work. I'm looking for something more automatable

[terse anvil]

idk ur current workflow, but i saw some workflows where they have a basic prompt and then make multiple prompts behind it to get multiple different results
you could also try to use controlnets with it if that is applicable in your use case, inpainting is nice for local fixes, if you really need regional prompting for some reason idk at the moment how to solve it

[terse anvil]

really depends on if its automatable in another way like my message above, if u really need regional prompting or can solve it another way

[terse anvil]

nobody said it was easy, also would help if you have a website or plan somewhere so people understand how your service works, because it sounds interesting but i have no idea what you are trying to get from people and what you will do for them in both benefit

[terse anvil]

a google search also make this pop out, might try it out: https://www.reddit.com/r/comfyui/comments/16qdpxn/regional_prompt_latent_composite_for_sdxl/

[wary maple]

That makes sense. I just posted a post in "community projects". In short, I'm looking to market their service on commission. I've had an explosion of demand for this type of service, so I see it as a win win

[terse anvil]

well if you are offered money for ai generated content, but need someone to generate it, it does sound interesting, ill hit you a dm ^^

[fallow veldt]

I use Automatic1111, and trying running OpenPose on top of regional prompter, getting the character in the exact spot/mask, etc, etc. OpenPose by itself is another offender, and with Regional Prompter, just a pile of garbage

[fallow veldt]

I tried Latent with Regional Prompter in Automatic1111, sometimes it works better, but takes much longer and it is also unreliable

It is like, as I said it is more easy to get multiple same characters with regular prompting (or maybe different characters) than multiple character with Regional Prompter. I tried many times and posted here, it is just a mess, and ControlNet doesn't help it.

It is like "A girl AND a boy" and it just make an single androgynous character in the middle, even with Latent

[terse anvil]

and openpose of 2 characters dont help? have you tried with weights, also other controlnets of 2 characters should also help, maybe play with weights, increase controlnet weights

[fallow veldt]

I tried, I would say it works better, but the pose usually is a mess, or characters appear any where, mixed characteristic, just unreliable

[terse anvil]

i work with comfyui and the control you have there is better imo, because you can easier change the way you want the workflow to act, so idk how to optimize it with auto1111

but my experience is that playing around usually helps to get the wanted result^^

[fallow veldt]

yes maybe Comfy has some more deep workflows than Automatic, but I would bet that in the end they work in a similar frustrating way xD

[terse anvil]

yes that is the way these kind of things work ^^

[fallow veldt]

I have learnt how to inpaint characters, it is some work but it works. The trick is to inpaint with fill and don't give it enough context so the other character is not visible (because if SD see it, it will think you are talking about it)

Problem is that as you don't give enough context, most of the time the character won't match the rest of the image. But when you get a base that make kind of sense, even if doesn't match much, you can inpaint with bigger context and it will get its attention on this new character.

[terse anvil]

afaik faceswap should only change the face

[hoary sapphire]

It generates like one Image the stops

I thought It was because of nsfw

[terse anvil]

probably setup the batching wrong

[hoary sapphire]

ohhhh, I shouldve looked at the terminal, sorry brother It was because of the resize resolution

[terse anvil]

np can happen ^^

[ember lily]

||help me anyone, i am unable to enter into bot for creating images, it shows, "You donot have permission..."😢||

[brittle heart]

Because it's OFFLINE

[sour needle]

Is there stable diffusion that works with mobile

I honestly don't think that's impossible

[warm junco]

Only on iPhone right Now

[normal fern]

Hello, quick question

Would you go for 4070SUPER (720€) or 7900XT ? (800€)

20GB Vram is enticing but heard amd is pretty bad with SD
(I also do play 4k on the same pc)

[full lark]

The 7900xt is better in games I think, but the 4070S is faster in all local AIs (llms or stable diffusion/svd) also maybe dlss will help closing the gap in some games

[wary zenith]

hello, could someone tell me how to have stable distribution on Kaggle

[fervent thunder]

google it

[oak tiger]

hello

here are saying

i am block from use the bots

what happened?

[tardy shard]

same

what happened?

[karmic brook]

#1047610792226340935

[karmic brook]

#1047610792226340935

[twin finch]

hi comfy when you will stream again

[unique sail]

O

[full lark]

Hello! Do anyone know if there is any way to make stable diffusion video generate faster, maybe using more vram or something?

[clear oyster]

its happening again, sdxl turbo images are all blurry like they needed more sample steps

sometinh about VAE? whats vae

[opal hedge]

Does anyone know if sdxl models that rely on clip skipping have difficulty with regional prompting

I know that's a bit of a general question, but for reference I am using comfy ui and ipadapter via RGB mask painting

[zenith citrus]

So I made a new lora :

https://civitai.com/models/288179

[opal hedge]

Damn, that's pretty freaking cool

Yoink

[zenith citrus]

thanks

[clear oyster]

how do I fix this

[forest trout]

Amusing. I like it!

[zenith citrus]

thank you

[clear oyster]

also my interrurpt and skip button dont come anmorem

[mild python]

Hello

[placid snow]

resources to get started with stable diffusion in google colab?

and using LoRA

[mortal echo]

anybody trying to make money generating images? hit me up!

[plush solstice]

Stability API is down and has been for a few hours, is it going to be fixed?

[iron depot]

So I'm following a ComfyUI tutorial and they have Load VAE nodes connected to VAE Decodes but then later they have the Load Checkpoint connected to VAE Decode and I don't get why or how that affects things

[golden gale]

For how long bots will be offline?

[bleak ice]

bots will be offline?

why can't paly the bot ?

[pale latch]

is there any truth to that reddit thread that boomed, where the guy claiming he proved the VAE is botched and he knows better than everyone ? He's one of these aiwars people and is going hard on trying to discredit stupid tools like nightshade, as if it even matters. they're obsessed and , it gives me cause for concern about his theory crafting.

[vague pond]

i mean unless someone else has a better explanation for what is going on, yeah, we're probably right about it, at this point there is plenty of evidence that something is amiss with the VAE, and by now we've actually observed some impact on training dynamics (namely that the high-mean, sensitive latent pixels do not change as much as others during training). we're not 100% sure on the details of the underlying mechanism, so a lot of that is subject to change.

[pale latch]

mm. so the obsession persists. ok

[tepid frost]

is it against the rules to try to hire someone from this community? paid, professional, corporate work

[vague pond]

and for what it's worth the lead researcher for DALL-E 3 did reply in there claiming that there are global information issues with the compvis kl-f8 encoder

considering how they created a substantially improved (if not horrendously memory hungry) improvement for the decoder, I would assume they have done their homework on it. they do claim that it's a great autoencoder nonetheless, and like I said, the models clearly do work. it's just that there's room for improvement.

could be training methods that mitigate the issue. could be repairing the vae. could be that this is something different but that's not likely

[balmy nebula]

Hello there! I just started my journey with stable diffusion and im impressed what can i achieved by using youtube tutorials! Is here a specific place where i can ask for help a people who know how it works and what can cause a problems that i met? ( #🤝｜tech-support i guess?)

[pallid canopy]

What the hell happened to the website?

It became 100 times worse

[hallow sable]

what website?

[pallid canopy]

https://stablediffusionweb.com/

Now you can only do 10 images for free, not to mention they cut off like 70% of the art styles for some reason

[hallow sable]

dont know never used it

[green raptor]

What is VAE?

[safe grail]

Excuse me , can i know why i can't use bot to do anything.....

[hallow sable]

its in the bot status

[warm junco]

This website has nothing to do with the official stable diffusion by Stability.ai.
The Stablediffusionweb.com site wants to trick users into thinking that its something official. Its not, its run by a shady company (black magic) to trick users, you even won't get your money back if you spend some.

[pallid canopy]

Whats the official site so i can generate something?

[warm junco]

Thats the official site but its paid only:
https://beta.dreamstudio.ai/generate

If you want something with free credits then best would be Leonardo.ai I guess

[pallid canopy]

Truly we can never escape capitalism

[warm junco]

You can also use the bots when they are back online for free here too

[twin finch]

can i have music bot please

in server

this server

[soft vault]

why do i not have permission to generate images

can anyone explain pls 😭

[teal pagoda]

Hello!
I want to start a team where we have people with some knowledge in the fields of generative artificial intelligence and marketing. The persons with bright ideas and an open mind for a collaboration can apply too.
We are doing this to lay the foundation for something unique that will also bring us an income. We will work 100% as a team and share the winnings equally.
Stop going alone, all the successful businesses have a strong team behind.
I have my own ideas, but I would also like to hear your ideas and opinions.
A little about me:
-> I am an AI Artist having an experience of over a year in the AI Generation field (using Stable Diffusion, MidJourney, Dall-E, AnimateDiff, FaceSwap etc.) and I can say that I am an expert right now, but I try to continuously learn new things/techniques.
Thank you for considering this!

Contact details:
-> Telegram: @Bullseye10001
-> Discord: bullseye3886
-> Email: [e-mail]

[pallid canopy]

What are you trying to achieve with this team

[teal pagoda]

Creating an unique service/product which can be delivered to customers.

Because believe it or not, AI is the future.

[knotty solstice]

Why don't i have permission to generate images on any of the bots?

[alpine nest]

Same

#bots show me that i don't have permission to send message. How to generate image?

[agile nacelle]

I use stable diffusion daily, But i've never looked for tools or other websites to help me.

I know of sites like hugginface / Civita AI etc. but are they are sites you use daily and recomment other creators to use.

[rose stirrup]

anyone knows of a free option for a1111 online? also if possible to upload my own models and loras

[thorn knoll]

why my gradio is loading forever

can someone help

[rose flower]

https://tenor.com/bk7l7.gif

[bold garden]

😆

[verbal pecan]

Hi

[hallow sable]

not only in AI but any program that does rendering or uses cuda/stream PU

blender with nivida cards render so fast that the 7900xt cant compete with a 4060

this goes with all programs. I haven't found a single program that a AMD out performs Nvidia. if someone is a pure game they should be buying AMD

[bleak matrix]

Good morning, everyone! How are we all this gorgeous morning?

[shell hazel]

even then gamers shouldn't buy AMD unless they need to save $ lol

[hallow sable]

why? amd cards are pretty good when it comes to games

[shell hazel]

Nvidia is king for 30 years

still is

lmao

AMD doesn't even have Frame generation

and the implimentation of it that they got recently is really bad

[steep star]

@west echo yo, are you still there?
been long since we talked

[wispy cloak]

where is free AI? (STXL)

[pale latch]

AMD has brought a lot of innovation as they've struggled to maintain the market share they've got. Without AMD , DX12 and Vulkan wouldn't exist. They, among a few other industry players who partnered with them, made the zero driver overhead model with Mantle. These new apis are objectively better. I don't think the industry would've moved towards their research if AMD didn't go all in on mantle.

competition is a good thing. don't be a corporate fanboy. Just enjoy the game as they play it. It's not like they're paying you to market for them.

[wraith cedar]

https://tenor.com/view/facepalm-really-stressed-mad-angry-gif-16109475

[static cape]

Does anyone know what @wise stratus was hinting at recently? Just hype creation or something to actually get excited for?

[pale latch]

dont believe emad's hype. it always falls through. when Joe starts talking, take notes

[static cape]

Yeah. It felt cheap and over the top... as if he was a Starfield dev. ^^

[pale latch]

Naw. Starfield landed exactly how they hyped it up. Skyrim in space.

More like Hello Games NMS devs

[static cape]

Hmm. I didn't buy into the Starfield hype from day 1. Turns out it was as boring as I had thought.

[pale latch]

I think it's another bethesda game. Its exactly what I expected. Its nice to have more bethesda fodder in the mix i think. between elder scrolls and fallout, they could use more IP

[real marten]

hey

i've been kinda living under a rock

where do I check new cool tools to use after uh.... the release of SDXL?

is autom11111 still a thing?

ty

[strong mulch]

What is the best training script (cli or something that I can automate) out there for training LORAs?

[full lark]

AMD is better in games for the price, but Nvidia is better in AI/ML and now more games use that (dlss/frame generation) AMD will compete when they come out with some DLSS equivalent

[silk cipher]

Hey, i'm looking for someone to help me develop custom stable diffusion pipelines using the diffusers library. It's a paid gig. If you know how, and are interested, plz DM me.

[rich kestrel]

How to view tensorboard on A1111

[vestal dew]

doorknell...as in "dead as a"

[nova zodiac]

Use something like a1111 or comfyui with the —api switch enabled…

[nova zodiac]

Kohya_ss or onetrainer are best trainers

[nova zodiac]

A1111 is still a thing, sdxl is cool, youll want to check out civitai to see the latest finetunes that people have made

Controlnet has added new modes that are cool, and the best update is probably animatediff/stable video diffusion

[strong mulch]

I just reinstalled Kohya to check it again. Its cli is works well enough for automation. I will check onetrainer for comparison

[warm junco]

AMD has FSR as DLSS equivalent and AFMF as Frame Generation equivalent to nvidia frame gen

[warm junco]

There is also this:
https://github.com/Akegarasu/lora-scripts

[main mica]

Hey friends, how many GB of RAM is recommended to be able to load an XL model? Every time I try, the model doesn't load.

Or is this not a RAM issue?

[rich kestrel]

so essentially you want people to jump through hoops and jury rig their setups just to do something that works right out of the box with nvidia cards. Ok👍

#teamgreen4life

[trim magnet]

theres no such thing as jumpin through hoops almost every modern game supports both FSR and DLSS

[stable yoke]

Hi everyone, I just arrived on Discord, and I don't have the ability to post in the 'bot-x' channels. Do I need to take a specific action?

[torn flame]

What model are the bots using in the bot channels from generation?

[nova zodiac]

8gb recommended minimum

[nova zodiac]

Sdxl 1 iirc

[main mica]

that's weird, i have 16GB and the model don't load

[main mica]

that's something that i have to do to use XL models?

[clear oyster]

I use a GPU with 6gb ram and there is about a GB of spillover into ram

its more the vram then the ram

sdxl turbo*

[main mica]

so there's any tip for me for how to start using de XL models?

I get a little frustrated because I downloaded the models, but when I try to load them on the web UI, it gives an error.

sometimes the webui crashes or just don't load

[trim magnet]

the more ram the better for XL if u dont have enough ram and also not enough vram the model will spillover into the pagefile in disk which will make the whole thing a slugfest

[cold shard]

https://www.youtube.com/watch?v=9k-yb83ZHfc

i have had luck with this video today i am still very new

[hard ruin]

Saludos amigos. Porque no puedo generar imagenes? Gracias

[main mica]

i'll check, thank you

this tutorial is for comfyui, i use webui

just for add more info to my issue: sometimes gives me an error of "failed to allocate memory" or something like that

or as said before, it just crashes

[cold shard]

sorry bro idk the difference ven

[main mica]

don't worry, i am new in these stuff too 🙂

[trim magnet]

whats your gpu

[main mica]

2060

this is what says

RuntimeError: [enforce fail at ..\c10\core\impl\alloc_cpu.cpp:72] data. DefaultCPUAllocator: not enough memory: you tried to allocate 52428800 bytes.

[hard ruin]

Hi everyone, I just arrived on Discord, and I don't have the ability to post in the 'bot-x' channels. Do I need to take a specific action?

[trim magnet]

you need to read the #1047610792226340935

[clear oyster]

with ddr5 ram I dont feel a significant performnc decrease with vram spillover

[clear oyster]

you need more ram

[trim magnet]

yea cuz its very fast,i have 16gb and runs very slow because i have lots of programs open that eat like 10gb of ram so only have like 5gb left

[clear oyster]

i have 16gb too

[clear oyster]

ddr4 or 5?

i fel like ddr5 is a lot better for the GPU

[main mica]

that's sad

[trim magnet]

ddr4

[main mica]

i have 16gb

I thought it might be enough

[trim magnet]

try with comfy,its better optimized for XL compared to webui

[main mica]

got it

i will try then

thank you

[hard ruin]

Greetings, they talk about migration. They're done but they're not ready yet. Could you explain to me what it is about? thank you very much and sorry for the audacity

[clear oyster]

16 should be fine, try closing other apps

[clear oyster]

im fine with 6gb vram and 16gb ram with automatic a111
no idea why it seems like im the only one not having a problem

oh you guys using xl ad not turbo?

[main mica]

i've tried

[trim magnet]

yea the normal one

[clear oyster]

oh thats proably it
im using turbo

thought that turbo would be worse thoh

[main mica]

models like dreamshaper 8 or Realistic vision works perfectly

but when i tried to load something like dreamshaper XL

nothing happen

just crashes

[clear oyster]

however yeah i guess xl needs more ram as it is spilling over

[main mica]

do you guys uses comfyui or webui?

or none of these

[clear oyster]

im using the a111 webui

comfy is supposedly better for xl though

[main mica]

I saw that ComfyUi is better in some aspects, but it seems so difficult to get started

i'm so sad i just wanted to use XL lol

life is cruel

[trim magnet]

use a turbo XL model it runs better because its smaller

[main mica]

okay, i'll try and see how it goes

[cold shard]

dont feel bad, i got comfyui fully installed and now i feel even worse because i have no clue how to operate it

[nova zodiac]

Ram or vram??

[fervent thunder]

how do i update the embedded python version? mine is 3.11.6 but SVD says it needs 3.11.7?

[nova zodiac]

Where is it embedded?

[fervent thunder]

within comfyui

[nova zodiac]

Local install or colab?

[fervent thunder]

local

[nova zodiac]

Stop comfy, Install 3.11.7 system wide, delete the venv and then restart comfy and it should rebuild with the new version

[fervent thunder]

thank you

[west echo]

I’m lurking 🙂

[steep star]

the SD server has been pretty inactive :sadge

is it the end of SD community?

[west echo]

I think a lot of people have moved over to AI revolution, theres a very active ComfyUI community there with a few of the node decs getting iinvolved

[steep star]

is AI revolution a app or something?

[west echo]

no , Its Olivio Sarakis discord

sent you an invite

[steep star]

but still it barely reaches the activity on #✨｜sdxl going on during july 2023

july and august were the months where the community was very active

[west echo]

news announcements, always more traffic lol

[steep star]

it was like 100 messages per minute

[steep star]

but even before that, all the devs were pretty active, joe is kinda dead in the chat, his 3 three messages were in novemeber, september and august 27

[west echo]

sometimes it can be overactive lol

[stuck sun]

is there a way to generate a couple of different prompt sets sequentially when testing checkpoint merges in supermerger?

[steep star]

those days were great

[compact prawn]

whats up

i have a question

what is the best computer t buy below 1300 dollars for good image creation

and by good i mean fast so it takes below 30 seconds to create images

[warm junco]

You need to edit the webui-user.bat.
At the line COMMANDLINE_ARGS=
you add: --xformers --medvram --no-half-vae
Then save and relaunch the webui-user.bat and try again

[nova zodiac]

I mean if you are using LCM lora and creating an image in 512x512, even a 1080Ti will take less than 30 seconds to make an image 😛

Generally if you are looking for a graphics card for SD, the best bang for buck is the 3060 with 12gb of VRAM, and then a computer to go around it

[warm junco]

Only 3060 has 12gb

The ti has 8 xD

[nova zodiac]

I keep getting them the wrong way round, but I don't think I'm wrong about that card being the good value one tho

[warm junco]

Yea 3060 is the best price/vram card still

[steep star]

i would suggest increasing your budget by $200,
you can get an used 3090 for around $600

[oblique willow]

hi

Is there any feature of stable diffusion which can select subject out of a image

Like,,remove background automatically

[west echo]

It’s the GPU that’s the main factor. That said my sub £200 1080 Ti can generate 1024x1024 images in under 30 seconds if I use LCM

[warped kiln]

hi i have a question should i use support ticket for that?

[cerulean ravine]

me too

[warped kiln]

hmm

[cerulean ravine]

I want to change people's facial expressions and I found some kind of wildcards thingy and I tried instsalling dyanimic prompts but it doesnt show up on the stable diffusion interface, I also installed wildcards and it also doesnt show up please help

[warm junco]

There is an extension for that called Rembg

[nova zodiac]

Are you using it on runpod or on your own machine?? If runpod, go into the jupyter notebook, go into stable-diffusion-webui folder and open the webui-user.sh file and under command line arguments delete —skip-install, then save and restart the pod

[rough burrow]

[cerulean ravine]

im using it on my own machine

[warped kiln]

when running the .bat file when downloading torch 2.0.1 cmd always fails to download so i downloaded it manualy the download is finished but idont know where to cut it to

do i have to download it from cmd

[jolly briar]

Díganme, ya no podré generar imágenes?

[warped kiln]

having to redownload a file on my computer that refuses to finish is killing me

[marsh lion]

Hello, it is possible to install Stable Diffusion on a server (Xeon, VMware, Windows Server 2022, lots of memory)? Actually trying...

[still glacier]

possible ? probably yes. Will you get good performance considering you're talking about xeon processors and not even mentioning gpu.... probably not.

[still glacier]

what's the error ? (also #🤝｜tech-support)

[still glacier]

#1047610792226340935

[warped kiln]

My internet connection is the error

Nothing larger than 50mb can download in 1 try

[normal warren]

I want to use the bot

[still glacier]

if you're using wifi, switch to a wired connection. Plus every model is about a couple of gb so you're gonna run into troubles later on anyways. Better try to fix your connection.

Very vague question.

[marsh lion]

Work! thanks to the --lowvram --precision full --no-half --skip-torch-cuda-test parameter 😉

[lilac forge]

I'm following Olivio Sarikas's video called InstantID for Automatic 1111. I have posted above, but I think my message got lost in the mix, so I'm posting here as well. Hopefully someone knows the fix to my issue:
I'm working on integrating InstantID with Stable Diffusion Automatic1111 but hitting a snag. I followed a tutorial requiring an up-to-date ControlNet extension. My ControlNet version (ed85cd1e from https://github.com/lllyasviel/ControlNet, downloaded on 2023-09-09) differs from the tutorial's (v1.1.440, data from 2023-09-10).

I also have the sd-webui-controlnet extension from https://github.com/Mikubill/sd-webui-controlnet.git. So does he, but he specificaly is pointing me to update the Illysviel ControlNet extension. Despite updating, I'm stuck with ControlNet v1.1.313, not his v1.1.440 as in the tutorial.

Following further instructions, I downloaded, renamed, and placed specific models from https://github.com/Mikubill/sd-webui-controlnet/discussions/2589 into {A1111_root}/models/ControlNet. However, I can't select InstantID options in the webUI. I only have "control_instant_id_sdxl [c5c25a50]" in the Model dropdown and no InstantID-related options in Control Type or Preprocessor.

He says to click on the blue refresh icon to the right of the Model, or to exit out of and restart and relaunch a new instance of Automatic1111. I have done both. He says "the first time you are going to use this, this is going to download the preprocessor models for you, so it's going to take a little bit of time because it's about 200 to 300 megabytes in additional downloads".
I'm not sure how to cause this download process that he is referring to.
Attached is a screenshot showing the discrepancies. How can I update to the correct ControlNet version or trigger the preprocessor model download? Thanks for any help!

UPDATE: I managed to get ControlNet v1.1.440 installed into my RunPod JupyterLab. I did this by uninstalling the illysaviel ControlNet extension and then, in JupyterLab, going to my extensions folder and then opening up a terminal and entering git clone https://github.com/lllyasviel/ControlNet . This ended up also allowing me to use the Preprocessors of instant_id_face_embedding and instant_id_face_keypoints. All was good and well until I clicked Generate Image. I feel like it's not pulling from the face at all.

I've uploaded screenshots of my settings. The exact same settings from Olivio Sarikas's video InstantID for Automatic 1111.

What am I doing wrong?

Edit: I can't upload a screenshot. So here are my settings:

Stable Diffusion checkpoint: turboDiffusionXL_v112.safetensors [1524a9699f]
Positive Prompt: 70's style Anime image of a woman ((masterpiece, award winning)),

Negative Prompt: ugly,deformed,watermark,frame,photo,realistic,3d,

Sampling Method: DPM++ SDE Karras
Sampling Steps: 8
Hires.fix : Off
Refiner: Off
Width: 1024
Height: 1024
Batch Count: 1
Batch Size: 1
CFG Scale: 4
Seed: -1

ControlNet Unit 0 [Instant_ID]:
Enabled
Control Type: Instant_ID
Proprocessor: instant_id_face_embedding
Model: control_instant_id_sdxl [c5c25a50]
Control Weight: 0.5
Starting Control Step: 0
Ending Control Step: 1
Preprocessor Resolution: 1024
Control Mode: Balanced
Resize Mode: Crop and Resize
Presets: New Preset

ControlNet Unit 1 [Instant_ID]:
Enabled
Control Type: Instant_ID
Proprocessor: instant_id_face_keypoints
Model: control_instant_id_sdxl [c5c25a50]
Control Weight: 0.5
Starting Control Step: 0
Ending Control Step: 0.25
Preprocessor Resolution: 1024
Control Mode: Balanced
Resize Mode: Crop and Resize
Presets: New Preset

[humble onyx]

can you get banned from here?

[dire sleet]

Why is the download for SVD 1.1 requiring agreement that it can only be used for non-commercial use? I have a SD subscription so I shouldn't be forced to agree to those terms.

You shouldn't be making people agree to conflicting terms of use.

[sturdy crown]

hi @here

[worn tulip]

sup

[main mica]

with this my computer freezes when I try to load the model lol, Perhaps I just can't seem to use these models 😦

[haughty smelt]

whats the one thats like stablediffusion

but u can download

[jaunty iron]

Hello, I have a question, I want to create a character in a real background, how can I do this?

[warm junco]

You mean a local webui?

[haughty smelt]

no i got it

its stable diffsion gui

[warm junco]

Ah okay

I thought you mean automatic1111

[plain raptor]

QUESTION

does the new stablle diff release still work with old models

from the bygone era

or nu

or is it like sony playstation, where it was backwards compatible, til 3.1 and then u cant play old games on it

[opal hedge]

Hey guys. Does anyone have a reliable way to fix hands in comfyui?

I've already tried handdetailer and meshgraphormer, I'm willing to try any other nodes, workflows, and even manual inpainting if necessary

[junior harness]

How good is a laptop 4080 gpu when it comes to creating photos with fooocus?

[nova zodiac]

You need the new controlnet extension version for the InstantID option to come up. Delete and reinstall the extension and it should give you the new version

[nova zodiac]

Yep, mods do mods

[nova zodiac]

Generate the background, then inpaint in a the character

[sharp brook]

Done

[nova zodiac]

Should be pretty great. Good amount of vram and new architecture means should do a decent job. Not as good as desktop version but still perfectly servicable

[fervent thunder]

Neymar with the golden ball in his hand

[nova zodiac]

The new models dont work with old loras, but the software that runs the models (eg comfyui/a1111) is fully backward compatible

[pale latch]

should clarify that further since there are new 1.5 models.

The old architecture works only with loras and embeddings and controlnet made for that architecture.

The sdxl architecture works only with loras and embeddings and controlnet made for that sdxl architecture.

both architectures have active communities, since the old one requires a lot less memory and a lot of people still rely on that

The software is compatible with many different kinds of models. Models are just billions of numbers that affect the neural network calculations. Those numbers are arranged in different "architectures" according to how the initial base model was trained

[languid willow]

Im trying for the last 15 hours straight to make outpainting work in stbl diff but with minimal success. I tried every tutorial. Every piece of content and blog was viewed by me. My problems are: AI generates outpaint that looks pretty terrible. If you have ANY tips to make outpainting on real images (not generated by AI) work I would really appreciate it. (PS: Im so down bad that Im thinking about buying an AI that can expand images or photoshop )

[nova zodiac]

Have you tried the mosaic outpaint extension in a1111? General thoughts with outpainting are use krita/ps and use block colours to give the sd some idea of what to make, then do img2img on that

Other thoughts are using krita and an extension that connects krita to a1111/comfy to do the generative fill stuff

[languid willow]

Thanks for all the info. I think I did not try mosaic? Probably not sure though because I tried a lot of extensions and options. Thanks once again I will try everything u said.

[junior harness]

How about mobile 4070 or 4060? First what matters is the memory only?

[grand mist]

depends

[grand mist]

like memory is most needed thing to create best graphics

so probably yeah

but my advice is to buy 4060 bc it work better for games and programs

and with much memory

[junior harness]

its 8gb vs the 12 gb of 4080

[grand mist]

oh

so just buy 4080

what type of img u will be creating

[junior harness]

Just images with fooocus

[warm junco]

4080 has 16gb

4060 ti too but is slower

for sdxl models start at 12gb, dont buy a 8gb gpu for SD

[fervent thunder]

any substitutes to stable diffusion while its down?>

[hallow sable]

uhh my dreambooth test models hate me 😦

they dont want to load

Creating model from config: C:\ai\stable-diffusion-webui\models\Stable-diffusion\KikiType1\KikiType1_1100.yaml
changing setting sd_model_checkpoint to KikiType1\KikiType1_1100.safetensors [f9130a8d2a]: RuntimeError
Traceback (most recent call last):

[junior harness]

The laptop version has 12 GB of vram

[warm junco]

oh right xD

[hallow sable]

I went with the 4060TI 16gb

Nvidia software is terrible

I need to use 3 programs to change the settings for the card with AMD everything is simple with more options in 1 program

[plain raptor]

oh noice :3

[pale latch]

i like that nvidia keeps the driver control panel the same all these years, and then the extras are in an app i can disregard, since other apps use those extra features better. AMD's changes constantly and attempts to hide complexity behind abstractions. Both have weaknesses. I think AMD's is prettier.

[hallow sable]

amd has way more options, so there is no behind being abstractions. if I want to access gpu fan curve I can just go to that setting, I still dont know how to set up freesync or limit fps on the Nvidia software

[pseudo bough]

anyone know a good sdxl lora or sdxl model that can handle stuff like lighting a cigarette or a lit cigarette in mouth?

[nova zodiac]

No decent one for sdxl on civitai, I would wager that it doesnt exist yet

[pseudo bough]

also can't get holding car key in ignition up close to work either

at least on the model I"m using

[pseudo bough]

thanks for the response though

[kind musk]

Has anyone fiddled with style aligned yet?
Running into a weird issue i wonder if anyone else has seen

[nova zodiac]

Yeah that sounds real tricky. Im guessing best bet is using photoshop/krita to paint in what you are after then img2img it

[pseudo bough]

i can probably use lyracanny (fooocus controlnet) to get it to work

[fervent thunder]

is it possible to invite the bot in our servers ?

[marble hawk]

I know Mage.Space gives unlimited free generations per day and Playground.com is 500 images per day

Are there any other good free ones to use until the Stable diffusion discord bots are back??

[nova zodiac]

Nope

[nova zodiac]

Stable Horde, HappyAccidents.ai, Leonardo

[marble hawk]

Thanks, I will look at those too

[fervent thunder]

Ah ok, thanks

[pale tartan]

About to download SDXL. I have a 2080ti, do I need something newer like a 4080 to generate images locally in a reasonable amount of time?

[nova zodiac]

2080ti should be pretty good. Using LCM methods will make it quite quick

[warm junco]

You need --xformers --medvram-sdxl --no-half-vae
In your webui-user.bat at the line COMMANDLINE_ARGS=

To get decent performance with sdxl models

[prisma agate]

Hi, im trying out stable diffusion with a 4090, but im only getting 15 iterations/s, what could I do to troubleshoot the low speed?

[nova zodiac]

15it/s at 512x512 or at 1024x1024??

[prisma agate]

512x768

Hmm, is downgrading to 531 drivers worth?

[nova zodiac]

That is on the low side - iirc there used to be things specific to the 4090 done to get the full 30-40it/s at 512x512

[idle plinth]

hi guys, how can I make nagative prompt for telling stable-audio NOT doing something?

like dash dash no?

[nova zodiac]

Can you weight prompts in stable-audio like (techno:1.3). If so, make the weight a negative number

[idle plinth]

what is the maximum range of figure of weighting? -10 to +10?

[prisma agate]

That didnt change much : /

[nova zodiac]

Theoretically there isnt a limit, but if you go too far from 0 you will get weird results

[prisma agate]

Also, should I be looking at the top it/s, or the highres ones? Because it shows me 15 it/s to begin, then at the total it says like 3.8 it/s

[nova zodiac]

The first one

[prisma agate]

hmm yeah, its nowhere near 30-40 xD

[warm junco]

How long does an 1024x1024, 30 steps image take with sdxl?

Oh you have a 4090?

Then yea --medvram-sdxl won't help much

[prisma agate]

about 6 seconds

[warm junco]

Thats pretty good

[prisma agate]

Might be, just dont want to be leaving performence locked off the table you know x)

[warm junco]

For max it/s you would need to use an 1.5 based model on a 512x512 resolution with 30 steps on Euler A sampler

Thats how the online performance charts on toms hardware were made

[prisma agate]

yeah, those are the settings im using, still 15 it/s

[warm junco]

With 6 seconds on sdxl I think your fine. But you can also use Sdxl LCM models

[prisma agate]

oh I think that might have been a 1.5 model actually, dont have any sdxl ones yet

[warm junco]

Ah okay

And you get 15it/s on 512x512 ?

[prisma agate]

yep

ill download an sdxl to try

[warm junco]

What If you use 50 steps?

[prisma agate]

still 15 it/s

[nova zodiac]

Yeah, need to do 50-100 steps to get it “warmed up”

[warm junco]

Also make sure your webui is updated

But if you installed it recently your fine

[prisma agate]

even at 100 steps and a few batch count its still the same

Ive got the thing where it tries to pull from the repo when I launch it, so it should be up to date..

[warm junco]

You can scroll to the bottom in browser and check the version.
Should be 1.7.0 and xformers 0.0.20

[prisma agate]

thats the one

torch is 2.0.1+cu118

[nova zodiac]

Batch count and size both 1??

[prisma agate]

yeah, tried a higher batch count just to "get it warmed up" but even at one its still the same

[nova zodiac]

Are you using xformers??

[prisma agate]

well, I do have that setting in the commandline arguments

idk if its actually being applied

[nova zodiac]

Look into 4090 cudnn stable diffusion as I think there were files that need to be swapped out to get full speed out of it

[prisma agate]

ill look into it, thanks : D

[short tapir]

does anyone have a link to the most up to date stable diffiusion xl?

i googled looking and the first few that come up seem super sus on websites cramemd with adds and stuff that make it seem like malware

[slate dust]

https://stablediffusion-online.com/stablediffusionxl.php

https://stablediffusionweb.com/

[short tapir]

uhhh how do i download that?

[slate dust]

oh sorry I thought you wanted an online version

Just download the models from Huggingface or Civitai into your fav UI https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0

[short tapir]

civati does models now??

[slate dust]

...that's like all it does

[short tapir]

like stable difusion? not just checkpoints?

[slate dust]

checkpoints are models

[short tapir]

like im trying to get the newest stable diffusion.. not models for it

like i uninstalled the stable diffuion i had becuase it was really old.. and am trying to find the latest up to date one todownload

[slate dust]

You just upgrade your installation.. don't uninstall it? Were you using WebUI?

[short tapir]

yes? and it wouldnt work with any models that used xl

[slate dust]

Then just install the newest version of WebUI

all models work

[short tapir]

i.. already uninstalled it all

[slate dust]

sucks

[short tapir]

thus why i asked for the download link to the latest

[slate dust]

an easy google search 🙂 https://github.com/AUTOMATIC1111/stable-diffusion-webui

[short tapir]

how do i even download off of hugging face?

[short tapir]

but is that just the ui? or the actual stable diffusion?

sorry super confused..

[slate dust]

Stable Diffusion IS whatever model you choose to use with whatever UI you choose to use. It's all the same thing

install WebUI then load models - easy

[short tapir]

but then how come when i had it before any model i out on that uses xl it just said moddel not loaded and refused to work?

[untold sleet]

can we make images again with AI in dream?

[slate dust]

Maybe you need some launch parameters depending on your machine specs. You can ask Tech Support

[slate dust]

#1047610792226340935 unfortunately not yet

[tired kindle]

What cloud GPU do people recommend?

[nova zodiac]

Depends how much time you want to spend on it. There is Paperspace which has a $8/month plan, but if you arent gonna spend ~30 hours a month making images, runpod and vast.ai are cheaper

There is also thinkdiffusion is well

But that is pricier. Depends how technically capable you are as paperspace requires a little bit of knowledge to get it setup, whereas runpod and vast.ai are easier, and thinkdiffusion it does it all for you

[warped kiln]

RuntimeError: Torch is not able to use GPU; add --skip-torch-cuda-test to COMMANDLINE_ARGS variable to disable this check

if i do as it says will it use my cpu instead?

[haughty brook]

https://takk8is.medium.com/beyond-consciousness-in-large-language-models-an-investigation-into-the-existence-of-a-soul-in-83d32002c3a0

[hallow sable]

yes

if you have AMD gpu you have to get the ver of SD that works with it

[umbral parcel]

So Stable diffusion is just down huh?

[warped kiln]

i found a potinatol solution which says uninstall and reinstall torch becaus of cloud flare downloads resault in this error i have a problem now that ichanged to google dns and deleted all files now it doesnt download from scratch it says Using cached data

[nova zodiac]

You will need to find and delete the python package cache

[thorny trellis]

What about to use cloud gpu to provide sdxl fine-tuning api

[warped kiln]

Thats it i give up

9 hours this stupid thing wont install

I downloaded fraking torch 4 times already

Now i can’t figure out how to clear python cache even uninstalling it wont do a thing

Only works when genrating on cpu and thats 8 hours ago gpu rendering is not happening

[hallow sable]

what gpu do you have...

[trim magnet]

skull issue

[nova zodiac]

All of the above will work for that

[nova zodiac]

Not helpful

[warped kiln]

rx 6800

[hallow sable]

did you grab the fork of stable diffusion / auto111111 that works with amd cards?

[warped kiln]

He’s right

[warped kiln]

I even tried amd offical method

Using ana or miniconda

I blame cloud flare

the amount of pain it caused me here in iraq

[hallow sable]

does the console say something failed? you might be able to download it by forcing a cmd with string in the batch file

[warped kiln]

Ill try that after deleting the cache

Thanks for the suggestion i will report back after redownloading

I belive i located the cache

Im confused why running webui.bat in admin immediately launches and shuts down cmd

[hallow sable]

dont run with admin

no need

[warped kiln]

I tried installing in a place with admin privileges but this happens so now i will just download with normal privileges

Ok

Hopefully the dns was the problem

[hallow sable]

should launch a tab in the deafult browser

[warped kiln]

I hope

Im getting linux flashbacks even tho if i tried this on linux it would probably install better

[primal nexus]

Hello, I have a question, when I create a sequence of softedge or openpose images from a video or just extract the frames of a video in ComfyUI, is it normal that saving the files to disk takes very long? like it saves only like 2 or 3 files each second so 300 frames take like more than a minute to save

combining them to a video just takes a fraction of the time

[warped kiln]

Didnt work slepp timeme

[viscid stag]

how do I make an api with sketch to image feature?

like I have my own data to train but I'm not sure how to do that

controlnet/t2i adapter?

like I wanna do what https://www.vizcom.ai/ did

[gaunt ruin]

/萨达达

[warm junco]

Hey, pls follow my install guide for AMD. You find it in the pinned messages of #🤝｜tech-support
It also covers the steps to clear the python cache.
So it will work 100%
For any questions feel free to ask.

[tacit elm]

Hello @drowsy meadow have you got an answer to your question about the possibility to use Coral Tpu accelerator instead of GPU? I am also interested by this possibility? 😉

[nova zodiac]

I believe there is a diffusers library that has a pipeline specific to IPU/TPUs

[placid hatch]

Anyone know the name of the automatic1111 webui extension that shows what parts of an image are influenced by particular parts of a prompt?

[brittle heart]

Stable it's not back yet? Did they forget about it? 🥴🥲

[lusty beacon]

What is the correct version of sd 2.1 and is it worth?

[warm junco]

its not worth

better use 1.5 models or sdxl

[hardy heath]

How do I make my training embedding faster

[plush reef]

hello

[hardy heath]

I have 8gb of ram

It's always stuck in 0 - 1%

Then after that shows error

Estimated time is showing 70hrs

[stark cobalt]

hello

we can't use stable diffusion rn?

[still glacier]

You can't use the discord bot. You can use any other SD solutions.

[short tapir]

trying to figure out what the extension is called. but theres one for promt where you start typing and it shows suggested promts and stuff

[lusty beacon]

thx

[warm junco]

boorutag autocomplete

is the name

[lusty beacon]

is invoke ai good or should i use anythig other

[short tapir]

ty

[lusty beacon]

i tried it for a while buti cant install it

[warm junco]

i would suggest to start with auto1111 for getting into SD

its the most used here

[lusty beacon]

okay

[warm junco]

you can follow my install guide, its in the Pinned Messages of #🤝｜tech-support

[short tapir]

uhg.. now the loras arnt showing up..

like they are in the right file and everything

[warm junco]

switch the model to 1.5 and they will show up

loras ar either made for 1.5 or sdxl

[short tapir]

ah.. so a lora i commsioned while using 1.5 wont woth with sdxl?

[warm junco]

a 1.5 lora wont work with sdxl right

[short tapir]

https://tenor.com/view/starwars-outrageous-unfair-gif-10552482

so then i cant make a waifu with xl then q_Q

[warm junco]

you could by using 1.5 to create the image, then load it to img2img, then switch to sdxl model and recreate the image while playing with the denois value

[tame thorn]

Hello everyone. How to use stable diffusion on this Discord channel, if it's possible?

[still glacier]

#1047610792226340935

[tame thorn]

:(

no one knows, when will it be back?

So, only option is downloading the model, and using it from device?

[still glacier]

or using other online solutions

(beware of scam websites)

[short tapir]

hm..

[ornate sage]

Hiii kujji

[short tapir]

so with my loras before i could put like different strenght sbut now when i put lora it dosnt work?

like when i add a lora it puts lora:rwby_cinder:1> and not a strenght?

[zealous hamlet]

I'm pissed the bot is still down. Like what the f the devs been doing a week now that they can't get it online? SmFh!!!

[finite plinth]

Ok

[still glacier]

<lora:rwby_cinder:1> I see the strength right there tho .... :1

[short tapir]

ah.

before when id put them it would auto put like 1.0

so i just thought the 1 was lora version not strength

[still glacier]

nope, it's the strength

[short tapir]

and idk what im doing wrong but before i could make good anime images, but now after reinstal im using the same model and loras as before but they seem... blurry and low detailed?

like the faces are all messed up now and it just.. dosnt look as good as before

[still glacier]

can't help without any more details, plus it's not the right place for that. It would probably be best to ask in #📝｜prompting-help or #🤝｜tech-support

[woven owl]

can automatic1111 do everything Fooocus can do?

[drowsy meadow]

hey @tacit elm, not that I know of, I followed along for the channel for a day or two, got busy after that unfortunately

I might just get one based on the return policy and return it if it doesnt work.
I wonder if there needs to be support from the ui tools that we use (comfyui and a1111)

also I'm currently struggling to do i2v on m3 Mac 32gb "ram".
just runs out of memory and machine just shuts down 😢

[lusty beacon]

anyone can help why my images dont look good

[jagged parrot]

Hi

[coral mist]

hi why the bot don't work anymore ?

[stark night]

ah, i asked in the wrong place, my apologies

[short tapir]

uhg.. got someone dming me trying to get me to join there ai server.. and they only found me becuase im a member here

[hallow sable]

where do I find a nonema model ?

[woven owl]

I installed reactor but somehow it's not visible in automatic1111

[fervent thunder]

If i generate images with batch count of 100 where I have added seed, should not alle the 100 images be same?

Becaus when I do batches each images are differenent.

[hasty iron]

hi

[twilit ginkgo]

Anybody here have experience making NFT's with SD?

[pale latch]

NFT images only have value because they're tied to criminal networks or the artist making them is already popular. The token itself is just a hash value, but it links to the image.

You can make an NFT that points at any kind of media in existence. An image made with SD is just an image. Wouldn't be any different

[twilit ginkgo]

we have an established project, and will be making NFT's for the community

but thanks for your info

[pale latch]

ahhh so you're looking for suckers for your ponzi scheme? makes sense

/bitconneeeeeeeect.gif

[sturdy briar]

Does anyone know why the bot is not working?

[twilit ginkgo]

cool opinion, thanks for taking the time out of your day to type about your opinion no-one cares about

[pale latch]

You'll find a lot of people will care about this fact as you try to sell nft access or whatever to your ponzi scheme. Don't be surprised when it happens. I fucking todaso Julian.

[twilit ginkgo]

they are free NFT's you bafoon

[pale latch]

uh huh uh huh. ok then. YOu're 5 years too late to the NFT hype train.

[twilit ginkgo]

another opinion that means nothing

thanks bro

[pale latch]

That scam pool has dried up

[golden gale]

is there any update about when bots will active again?

[hallow sable]

imaging being such a pos that instead of getting a job you try to scam people with nft lmao

[nova zodiac]

#1047610792226340935 is all the info we have currently

[stuck sun]

is there a good tutorial on merging loras into checkpoints without mucking them up (in supermerger)

[pseudo lark]

I made a customGPT in chatGPT that gives good and good length descriptions. Try it out, if you have any suggestions of what can be better let me know 🙂

https://chat.openai.com/g/g-EQFzMkKHZ-image-descriptor

You can upload multiple images at once, you don’t need to add any text. It will just give you a zip file with the descriptions in the txt files 🙂
I think if we try to make a big dataset out of this to make a better captioner we could get SD to the level of dalle3
(You need chatGPTplus to use this) let me know if you want to help compile a good dataset.
You can caption up to 400 images every 3hours with this due to chatgpt limits

[stray plank]

Wonder if it will work with SD?
https://youtu.be/wxlHq_EuxFU?si=9C4y0p72nDOw-ZEt

[hollow gazelle]

Hi all, wondering if the bot only works on this server or whether it can be invited to another or DM'd to make generations?

[nova zodiac]

dunno... I know that once you've tested it and got the strength you change the weight in the bottom bit rather than changing the alpha

[nova zodiac]

It cannot... it also isn't working currently #1047610792226340935

[frigid sinew]

hello would it be possible to LoRa train an image generator on gore?
I can curate the dataset but I would prefer to not describe and place trigger word for each image?
I would like to make it so I can replace the face/person with someone I send a photo of
Can I pay a freelancer from a third world country for describing each image?
or is anyone interested in training the entire thing for me? I can pay a bit of money?

including curating and describing the images

but I can host the qlora training

[long wigeon]

Maybe a very nooby question, but is there a way to restart the webui on mac?

I've always run it through cloud solutions and they had a built in restart button

[hallow sable]

should be in the folder you put SD in... I dont know what file it is because mac is horrible OS and I stay far away from it

[long wigeon]

What should be in the folder?

[hallow sable]

in windows its a batch file called webui user

but on linux or mac its probably another script file

[long wigeon]

It's called the same, but how does that help me restart it? I thought there might be a terminal command or something

[teal pagoda]

Starting a Winning Team with Brilliant Minds

Hello!
I want to start a team where we have people with some knowledge in the fields of generative artificial intelligence and marketing/SEO/dev. The persons with bright ideas and an open mind for a collaboration can apply too.
We are doing this to lay the foundation for something unique that will also bring us an income. We will work 100% as a team and share the winnings equally.
Stop going alone, all the successful businesses have a strong team behind.
I have my own ideas, but I would also like to hear your ideas and opinions.
A little about me:
-> I am an AI Artist having an experience of over a year in the AI Generation field (using Stable Diffusion, MidJourney, Dall-E, AnimateDiff, FaceSwap etc.) and I can say that I am an expert right now, but I try to continuously learn new things/techniques.
Thank you for considering this!

Some of our ideas:

1. A chatbot combined with Stable Diffusion/MidJourney API which will talk with people who feel lonely, depressed and can send pictures to make them have a better mood.
2. An AI Image upscaler service like magnific.ai, but better at keeping the input's image details and not changing it very much while increasing the overall quality.
3. Creating AI influencers for top notch companies to advertise their jewelry/clothes.

Qualities we are looking for:

• punctuality
• seriousity
• trustworthy
• team player

Contact me for more details.

[nova zodiac]

CTRL + C ends the session, then run webui.sh again

[sick flame]

Hello guys, I have a question. I hope someone can help me here. How can I train a model on my photos to be in a specific anime style? I have a data set for anime and I want to train my photos to this style. Is there a way?

[nova zodiac]

1. Dall-E in bing chat, chatgpt4, and Google Bard can all do this already...
2. thats just using magnific.ai but with a slightly higher denoise setting..
3. thats chatnexus.ai

Good luck to you though!

[nova zodiac]

Yup.. look on youtube for "lora training"

[sick flame]

I want to try what Corridor Crew did, but in a simple way and for a short clip. Would your suggestion be the most appropriate? and thank you also

[nova zodiac]

No idea what Corridor Crew did tbh

[sick flame]

May I share with you the YouTube link so you can see what I mean?

[nova zodiac]

as long as it is sfw

[sick flame]

https://youtu.be/GVT3WUa-48Y?si=GDPJ-zR3fG6sLhWB

[sick flame]

Yes offcourse

[mortal delta]

i dont even think youtube permits naughty stuff.

[nova zodiac]

yeah, you'll want to make the lora then

[sick flame]

Thank you for your advice. I will inform you of the results later

[mortal delta]

can i ask questions in this channel or are there better channels to do so at?

[fervent thunder]

hey everyone, I am bored of midjourney and looking to get into SD. I am not a gamer at all but looking at video cards. Would I regret only getting something like a RTX 4070 or would going much higher be overkill at this point when I have no idea what I am doing? ComfyUI looks so amazing so I am looking to get into that for sure.

[mortal delta]

The minimal vram you need is 4 bu ti would recommend getting a card with atleast 8vram, but im no expert so take my words with a grain of salt.

[fervent thunder]

Thanks. Yea I have no reference for any of this. My card right now is like the minimum that can drive a modern monitor.

[nova zodiac]

Try using runpod or vast.ai first (if you don't have a decent gpu already).. if you like it from then, you can invest

[mortal delta]

so how much vram does your card have.

[mortal delta]

doesent that cost money?

[nova zodiac]

fuck all compared to buying a 4070 and realising you don't really like it and prefer mj

[fervent thunder]

I will check that out. I can't imagine not liking comfyUI though. It is really video that I want to get into eventually.

[mortal delta]

Im very annoyed i got a graphics card but it cant fit into any of the computers i have, now i have to buy a motherboard too.

[nova zodiac]

I looked at the math and realised where I am I would need to use ~4000 hours of runpod to break even with a gpu, and I can't justify 2.5 hours a day for the next 5-6 years 😛

[fervent thunder]

I really have no idea what my video card even is right now. I need to build a new rig anyway. I am running this ancient i5 at this point so I want to get a modern video card. The cost isn't a huge issue if it is good bang for the buck performance wise.

[mortal delta]

well i guess good for you?

anyways my current pc has 2vram and i really want to try stable diffusion with what i have...

but idk

[fervent thunder]

Ah that is good to know. I will enjoy researching this.

[sturdy crown]

gm

[mortal delta]

its like almost nightime here.

[nova zodiac]

its more to point out that if you haven't got a gpu already, getting one solely for SD might not be a good value proposition

[nova zodiac]

its 2pm tomorrow here lol 😛

[fervent thunder]

Hold on, I use kde neon. Let me see if it shows the device. It must be a shitty video card I have from 10 years ago.

[mortal delta]

i have a 2vram one, and an 8vram one that wont fit in my pc.

im currently using the 2vram one for low end gaming.

but im impatient and i want to use sd locally.

im not sure if i can even run stable diffsion.

[fervent thunder]

Oh yea my video card burned out a few years ago. I forgot about that. I have
Intel Corporation HD Graphics 530 (rev 06)

[mortal delta]

that isint very great to say.

[fervent thunder]

Yea I am not a gamer but maybe I would be if I got a good video card. On the other hand I am not really sure I want to get that into games anyway. I will look at a cloud option for sure first.

[trim magnet]

i saw an ankle there,thats a big no no

[frigid sinew]

hello would it be possible to LoRa train an image generator on gore?
I can curate the dataset but I would prefer to not describe and place trigger word for each image?
I would like to make it so I can replace the face/person with someone I send a photo of
Can I pay a freelancer from a third world country for describing each image?
or is anyone interested in training the entire thing for me? I can pay a bit of money?
including curating and describing the images
but I can host the qlora training

[nova zodiac]

possible yes, dataset fine, face swap - use InstantID/Reactor/ipadapter after you have made the lora, captioning - get AI to do it using BLIP, outsourcing - not personally; try the bounty board on Civitai

[sturdy crown]

I can help you. @trim magnet

gm

Hi. could I have a chance of work with you? I am fashion designer as well as frontend dev @karmic brook @sudden ruin @bleak matrix @still glacier @austere marsh @gritty scarab @neon oriole @hidden dagger

[trim magnet]

sorry but help is not required

[bleak matrix]

Please do not mass ping people! If you are interested in jobs at Stability, you can check out https://stability.ai/careers

[fervent thunder]

That is pretty sick, I am pretty much up and going on runpod with comfyui that fast.

[fervent thunder]

civitai.com is all SD art? I don't know, if that is the case it would be tough to say SD is on the level of midjourney. Awhile back I remember someone on hacker news saying midjourney sucks if you know SD well.

[stuck sun]

MJ is a much much better general purpose model but you cant make it do really specific things

[nova zodiac]

if you know what you are doing, you can get the models on civitai to do far better things than are posted there (mostly model cards are there to give you a taste of what they can do, while still having the images reproducible by simple methods)

If you push the models using complex workflows and are prepared to spend some time inpainting, you will beat mj in most categories

and yeah, it can do certain things well, but for everything else, theres loras...

[frigid sinew]

hello

[frigid sinew]

nice pfp is that OC?

[vague pond]

you can do far more with SD than with midjourney due to the simple fact that you can train SD and also make custom changes to inference, instead of your only exposure to the model being a literal discord bot

[frigid sinew]

can u explain? @vague pond

I like midjourney cause the quality is really good

[pale latch]

use a riser cable

[gleaming totem]

so what would you guys suggest 3060 12gb or 3080 10gb?

[nova zodiac]

3060gb 12gb

cheaper, more vram, not a massive performance hit

[gleaming totem]

any general reasoning for the 3060, i saw an article that had benchmarks where the 3060 was ususally faster for 512 but when adding control net the 3060s performance tanked and the 3080 pulled ahead vastly

ty for responding btw 🙂

[nova zodiac]

more vram just allows you to do more stuff.. I can wait an extra minute if it means I can make cooler things

[gleaming totem]

would the 4060 ti be a great improvement over the 3060 / 3080 prices @ $480 : 4060ti 16gb $320 for the 3080 and $240 for the 3060. would the performance and vram increase warrant 2x the price and maybe slight future proofing? any caveats?

[sturdy crown]

My name is Peter Andersson I am fashion designer as well as I am frontend dev. Often using 3ds max, photoshop etc. I make web sites very fashionably. I wanna contribute by means of my skills. I speak in English. Thank you.

[nova zodiac]

the 4060ti isnt any faster than the 3060 due to the fact they hamstrung the memory bus

[gleaming totem]

https://www.msi.com/blog/stable-diffusion-xl-best-value-rtx-graphics-card#:~:text=Speed(sec.)&text=Following up second%2C thanks to,slow time of 65.1 seconds!

im going off of this article and i know it had 128 bus but if the benchmarks above are true it would actually be a bit faster.. im here with a rx 580 and i got it to work with directml abut i get oom errors and stopped tinkkering with it in hopes of finding an affordable upgrade

[nova zodiac]

Interesting - the addition of the controlnet must use more than 12 of vram and thus the 3060 had to start swapping in and out of system RAM

whereas the 3080 may have been doing that all along

[gleaming totem]

i think ur definately correct about that

[nova zodiac]

if you can justify the extra $$ then go for it

also, are those prices USD??

[gleaming totem]

yes

[lusty beacon]

How do people make AI Influencers with the same Person in different settings?

[nova zodiac]

Shit.. I can't get a 4060ti for less than $600 USD here 😦

[nova zodiac]

they have either trained a LoRA of that person, or they are using ipadaptor/reactor/instantID

[gleaming totem]

is there a big diff in manufactuere , like any to stay clear of

[gleaming totem]

best buy and amazon have them for around $450 atm

[lusty beacon]

But if this Person does not exist how can they train a lora

[nova zodiac]

$350 for the 3060 here

[nova zodiac]

make one image of the person, then train the lora off the single image

also some models are good at giving the same face for the same negative prompt

[gleaming totem]

i know a guy (lol) i can get the 3060 for 240 or so and a 3080 for 320 im trying to jump on it before its gone

[nova zodiac]

and there is also the multi celebrity trick as well where you combine 3-4 celebrity names to get a consistent face, then train the lora off those images

[lusty beacon]

mhh okay but In tried to make a lora from 9 images and it was a mess

[nova zodiac]

there are tricks to it - there is an article on Civitai on how to make a single image lora

but if you can use the multi celebrity trick to get out to 15-20 good images, then you are through to the semis without dropping a set

[trim magnet]

if u want 12gb and good speed just buy the 4070 12gb

[gleaming totem]

lol thank you , i was just about to click the buy button >.<

[trim magnet]

yea and i think the price dropped because of the super release

[lusty beacon]

it's complicated

[bleak matrix]

Good morning, everyone! How are we all this gorgeous day?

[lusty beacon]

Can I use Stable Diffusion Video in Invoke ai?

[acoustic vault]

Is Ai allowed in video games yet?

[hybrid marsh]

Hi all,
who here has been thinking of making his/her own foundational model?
be it text to text/image/video

I would love to talk with you....

[sharp meteor]

Has anyone tried printing their SD work out? I've got some questions regarding DPI

[vocal prism]

bots are downs for a while, will they be on again?

[sudden ruin]

They will, just no ETA as far as I know

[sudden ruin]

Yes, what questions you got?

[sudden ruin]

Why wouldnt it be? Just needs to be marked on steam from my knowledge

[sharp meteor]

Is there a way to set SD to use 300 DPI as default? Regardless of the format and resolution the images are always in 72DPI

[acoustic vault]

cuz last I saw it was not lol, steam allowed it recently

[sudden ruin]

Not that I know of, I convert them in Photoshop to 300 DPI

[sharp meteor]

And that doesn't screw with the image quality? I was under the impression that in order for the print to be ok it has to be saved in 300DPI. Can I do the same in say, Photopea or Canva for example? Looking for a free alternative

[sudden ruin]

Thing with DPI is it doesnt change the image as long as you dont try to keep the physical size the same, if you recalculate that it will obviously cause a reduction in quality. So if you got a 10x10 cm image and change the DPI from 72 to 300 it will only have around 2.5 cm left but with the higher DPI

[sudden ruin]

So it really depends on what you need, but there always is the option to upscale your image

[sharp meteor]

Well what I need is to keep the same size of the image but have it in 300DPI so it can be printed. I've already upscaled the image to 4k

[lusty beacon]

Hey is there an AI which can help me build prompts?

[sudden ruin]

Then you're bound to upscale further unfortunately. Thats my issue with AI as well currently, its meant for web and not really for print so far, at least not without some workarounds

[sharp meteor]

Guess I am, having a DPI slider in the UI will make our lives so much better. In time. Since we're talking... Can I bother you about one more thing? It's about tiling and transitions.

[tacit ridge]

hey everyone

[sharp meteor]

hi

[manic grove]

OMG bot is working again !

That's what I'll say when it does 😆 .

[mortal kraken]

hey hey

[cobalt sparrow]

Hey

[sudden ruin]

It doesnt make any sense having a dpi slider, cause it would just mean rendering the images at a higher pixel resolution. So we just have to wait for higher res models.

Whats the other question?

[sharp meteor]

So I'm trying to make seamless tiles but after zooming in the transitions are visible, is there a way to fix that?

I'll PM you a screenshot, can't post it here for some reason

[sudden ruin]

You can use #🏞｜general-with-images

Id probably also fix that in PS, there is a mode to edit tiling textures

Not sure why SD messes it up tho

[sharp meteor]

So there's no way to correct that in SD huh?

[sudden ruin]

I dont think so, but I also didnt use that feature a lot

[sharp meteor]

Gotcha. Well thanks for the help anyway, I've taken enough of your time.

[sudden ruin]

No worries, here to help

[plain raptor]

anyone mind helping me quick ;^;

i did sum, and i dunt remember wut i did to do it

for a friend

[grizzled mortar]

Hey all, im having issues with unclip. With the installation for automatic1111, am I correct to think that I just have to download the models and put them in the ../models/stablediffusion folder and that's it? Then I can just go ahead and use it in the img2img after I set it as the checkpoint? The issue im having is when I create an image via another image, it doesn't look anywhere close to what clipdrop is showing for the same image ( with their reimagine model).

[plain raptor]

<v> this is a--

<v>

[sudden ruin]

please move this to #🌶｜off-topic

[sudden ruin]

What help do you need

[minor bobcat]

k

[plain raptor]

tech support channel ;^;

[hybrid falcon]

why I cant generate images with the bots?

[mild grove]

guys, in comfy ui, suppose I have a KSAMPLER with input from another KSAMPLER, how do I play, only on the second ksampler, I always need to generate the first one again, the buffer is already loaded inside, but for it to go to the next node, I would need to play everything again , is there any solution?
my workflow depends on one latent coming from another latent, but playing them all costs me a lot of time

[sudden ruin]

Bots are down as stated in #1047610792226340935

[pallid sable]

wen up?

[nova zodiac]

Next time they have a new model to show off usually…

[nova zodiac]

Convert the first outputs to pixels and save it, then load the image back in in follow up gens

[clear oyster]

how much ram does stable diffusion video need

[scarlet shell]

Hey - I'm having a world of a time trying to get Kohya's GUI installed. For some reason, my python version isn't changing? I think that's the problem. I've downloaded 3.10.9, changed the environment varible path to 3.10, and I still get an error on the setup.bin. Any ideas of what I'm doing wrong?

[nova zodiac]

#🤝｜tech-support may have a better option.. you will need to restart the whole machine after loading.. and if you are using a venv you will need to delete and recreate that as well

[brisk cobalt]

If I gave someone an image, could they convert it to a prompt for me?

I dont have automatic 1111 installed rn

😢

[nova zodiac]

https://supagruen.github.io/StableDiffusion-CheatSheet/ - go to the metadata tab and upload image there

[brisk cobalt]

thanks soldier! 🙂

[mild ferry]

can someone teach me how to make a lora for a face made by AI. (I want this because when I use roop/faceswap instead, it gives me artificial results. for example, when using faceswapper, if i had "slight red blushing cheeks" in the prompt, it wont make a difference

[jade wedge]

Any advice on another bot or ia for video generation while the bot is offline? Thanks

[terse plume]

Hi what are you generating?

[charred wigeon]

Hello friends, I'm using SDXL with Fooocus, and I'm trying to figure out how to make it clear to the model that a specific name or person (for example, Metal Gear Ray from the Metal Gear Solid video games) is also understood as "Metal Gear Ray." This way, I can simply write "Metal Gear Ray staying in a desert" in my prompts, and it understands what Metal Gear Ray is. Can anybody help? thank you very much in advance

[nova zodiac]

Nothing.. that is an online version of the PNG info tool in a1111

[nova zodiac]

you have to train it using a lora

[charred wigeon]

can you provide a link?iam very thankful for your answer

[nova zodiac]

Online generators: Civitai, tensor.art, happyaccidents.ai, shakker.ai, tinybots.net/artbot

[jade wedge]

Thank you

[nova zodiac]

https://www.youtube.com/watch?v=70H03cv57-o - it's an older guide but it checks out

[charred wigeon]

thank you so much

[nova zodiac]

^^

[charred wigeon]

checking

[ripe kraken]

I generated a picture of a face and now I want to use that face and generate a body for it, how should I go about doing this?

[charred wigeon]

try to use stuff like angle, perspective, lens (35mm etc) and tell the model at the end of your prompt how the shooting angle should be

and of course explain what he/she should wear (for example: Oversized Coat in black)

[ripe kraken]

alright, I think I need better negatives because I'm getting 6 toes, extra fingers, etc.

[charred wigeon]

i dont know which model you are using, but yeah, we face all these "problems"

[marble steppe]

so, if I setup stable diffusion on a computer, is it allowed to let other people generate stuff with it and charge them a fee ?

[charred wigeon]

nice business idea haha, its open source, u can do whatever

[marble steppe]

oh.
Isn't that what's already happening? with websites that give you like 50 credits a month or whatever, then they offer plans

[ripe kraken]

yeah there's a lot of websites that do that

it seems very saturated

[charred wigeon]

yep, u are damn right

[marble steppe]

I thought so.....

What kind of hardware is needed in terms of motherboard/cpu/ram?
I have two 5700 XT's and an RX 6800 I could setup on a Linux machine....but it's a mining rig currently, so the mining motherboard iuseless for anything but.

[charred wigeon]

this should work fine

more then fine

no worries

[marble steppe]

right, but the motherboard is shit and only uses one PCIe lane per video card, has a built in mobile processor about as smart as a fucking toaster, and about 2gb of ram built in.

so I need to buy a new motherboard, processor, and ram

[charred wigeon]

yep

[marble steppe]

thats what I was asking about....

[charred wigeon]

gotchu, sry

[marble steppe]

suggestions that would work well for a stable diffusion machine

[charred wigeon]

bro, whatever mainboard around 100 bucks

will do the job

[marble steppe]

I haven't built a computer for a while aside from the mining rig, but like I said, it didn't need much lol

[charred wigeon]

most likely cheaper ones too

just run with commercial good brands, take the cheapest mainboard and u good to go

i have 6k machine running with a 200 bucks mainboard

[marble steppe]

probably need a ton of RAM though right? like 64+ GB?

[charred wigeon]

for 5 years

no fucking issue

[marble steppe]

or is it all about video memory?

[charred wigeon]

and its holding two 3090 TI´´s

[charred wigeon]

most likely vram, the clock speed is important too, but its backed up by the vram

more vram - right clock speeds

nobody cares

you dont need to be way to specific about your machine - stop melting your brain : D

your GPU´s will do the job - 100%

[nova zodiac]

All about the vram...

[marble steppe]

I was just curious about what other requirements were needed to run SD effectively. so it's especially good to know that I don't need (more) expensive ass hardware to set it up.

these video cards were fucking stupid expensive back before ETH switched to proof of stake.

[charred wigeon]

your RX 6800 will do a perfect job, as long you are not running a production company lol

[marble steppe]

I only wish IU'd stayed with nvidia, but AMD was better for mining

I'd still run all 3 if I'm able to

can I use all 3 together for SD?

[charred wigeon]

again stop melting your brain... start creating ❤️

[marble steppe]

nvidia is better for gaming! lol

and apparently better for SD on a windows machine? I'm going to run linux anyway

[charred wigeon]

https://www.reddit.com/r/StableDiffusion/comments/13omwgv/multi_gpu_support/

[marble steppe]

well thats even more interesting 🙂 actually kind of happy about that because it still leaves me two 5700 XT's lol

[charred wigeon]

By installing the RockEm and Stable Diffusion setup on Linux, users can enjoy up to eight times faster processing speeds compared to Windows.

[marble steppe]

fuck me running I might do something stupid and hit up canada computers tonight lol.

[charred wigeon]

but... depends on resolution, u wont see a difference on lets say, 4 created images - they will genreated with 16GBs, fast, anyways,

seconds

[marble steppe]

what do you mean?
I have upscaling tools anyway

[charred wigeon]

iam just saying. lets say you prompt a image, and u set the number of variatiosn to 4. so you creating 4 pictures. u wont see any huge time difference between linux or windows. lets say you use a prompt and tell the model to create 32 pictures. one after another. u would probably see a little time difference

[marble steppe]

oh ojk

I've got a lot of learning to do....what about training it for things? can you only train it with like one set of data or whatever or can you just keep giving it stuff....?

as if I wasn't clear that I have a lot to learn and don't know wtf I'm talking about lol

[charred wigeon]

you mean like, you have a specific charachter, thing, or whatever and the model doesnt understand what u mean? i asked a similar question couple of mins ago. you can install extensions and train the model specific stuff

thats possible

[forest osprey]

Yoo guys does anybody know when will the bot channels be available to generate images please ??

[charred wigeon]

iam installing it right now

[charred wigeon]

sry dont know, iam running the model locally

[forest osprey]

What do you mean by that

[marble steppe]

I've just heard about check points or whatever

[charred wigeon]

downloading it and running it on my pc

[forest osprey]

How?

[charred wigeon]

https://www.youtube.com/watch?v=70H03cv57-o

[charred wigeon]

one sec

posting u a tutorial

[charred wigeon]

https://www.youtube.com/watch?v=zIhODzEVZqg&t=49s

here you go

[forest osprey]

Got it boss thanks 🫡🧡

[charred wigeon]

go, create and have fun my friend

[forest osprey]

Sure brother 🥳🙏🏻🧡

[charred wigeon]

LFG

[fervent thunder]

Thanks. I am going to stick with it. There are so many aspects of midjourney I am tired of and just don't want to deal with anymore.

[nova zodiac]

I'm sorta surprised they've managed to keep going this long with just a discord bot and a recent beta web interface

[charred wigeon]

they are about to launch it on web - closed beta was or is running, i wasnt checking that again

[nova zodiac]

they do have fantastic models (both base and niji), and if they had the ability to do some of the advanced workflows that sd can (controlnet, ultimate upscale, animatediff/deforum), they would be nearly unstoppable

although then again, they keep it quite censored, so people hate that

[charred wigeon]

its all about the money, they could do all this, by now

there model is miles in front of any other model - but i know, there will be something way bigger on the way in the next few months

[fervent thunder]

Yea I guess I haven't minded it. I had never used discord previously and would never have. The community aspect is kind of cool. It is hard to say for sure with anything but I wonder if the compute is capped on Saturdays during peak times because I just can't understand how some prompts can give such poor results between sessions. It could just be totally random though.

For me I still like the 5.1 raw model the best.

[charred wigeon]

yeah, but its so fucking slow

"slow" compared to SDXL

i prompt 30 imagines in the time, midjourney is creating me 4 low resolution ones. if somebody got the "time" in this fast pace world.. ofcourse midjourney it is but otherwise its just to slow

[nova zodiac]

yeah, especially with LCM these days

[charred wigeon]

and all that cool features, midjourney has will be implemented by a crack chead for sdxl anyways

couple more months....

[charred wigeon]

whats hot today, is trash tomorrow

shit got ridiculous

ridicilous, in this fast pace world

[fervent thunder]

oh nice. I am just starting to learn. I am here because some of the videos I seen on the SD reddit group. Really amazing stuff.

Well I used 1.x whatever probably over a year ago when it wasn't so great then went to midjourney right when 2 came out.

[charred wigeon]

@nova zodiac thx again for the LORA tip, installing rn

[quaint saffron]

Hey guys, anyone using animatediff in here?

anyway to make it go faster? I use 22 second video.. rtx 3090, it is taking foreverrrrrrrrr

720p

[nova zodiac]

use the lcm lora?

[quaint saffron]

@nova zodiac what is that? I'm in comfy ui

[nova zodiac]

https://huggingface.co/docs/diffusers/main/en/using-diffusers/lcm

download the lora and use with the LCM sampler...

[quaint saffron]

does it work with sdxl? or standard sd only

[vale otter]

why sd-webui inpaint/sketch is so buggy... nobody actually uses it in reality?

[nova zodiac]

There are loras for both sdxl and 1.5

[nova zodiac]

inpaint is tricky - use a sdxl model, or use a 1.5 inpaint model and change the resolution to 768x768; then get used to masking, generating, clearing the old mask before making a new one, and then doing a final img2img to get rid of weird mask blurs

[gritty pelican]

is there any chat were i can share a video i just made to get feedback?

[vale otter]

that's not what I'm talking about

https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/8872

for example

[quaint saffron]

Thank you

[vale otter]

or when I send generated img to inpait, only a fragment of it appears in it...

[nova zodiac]

ah.. yeah.. your best bet if doing a lot of inpainting would be to install an extension that allows you to use it in Krita/Photoshop

[marble steppe]

What kind of HD space is needed for SD, and which Linux distro is best? 128GB good enough?

[nova zodiac]

HD - depends on how many models you want to keep, and how many extension/custom nodes that have additional models you want installed

[marble steppe]

those are words I understand individually

how many models?

[nova zodiac]

base install for a1111 is about 10gb, and then 2gb for each 1.5 model and 7gb for each sdxl model

then things like animatediff or controlnet have additional models that are a couple of gb each

[marble steppe]

what is a1111?
I have a lot to learn lol

[nova zodiac]

so if you only have 1-2 models that you use regularly, 128gb is fine, but if you like having lots of models available, then you'll want 256gb+

automatic1111 (also known as sd-webui) - it's a user interface that allows you to generate

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

[marble steppe]

and if I want to run it as a bot on discord?

[nova zodiac]

it has a --api switch

which exposes the api

[marble steppe]

I have no idea what that means in terms of discord bots, because I don't even know how to setup a discord bot or much about discord for that matter

at any rate, starting from the beginning, which linux distro is best for SDXL?

[nova zodiac]

are you just looking for a way to generate an image?

or are you trying to build something like midjourney for others to use?

[marble steppe]

I want to install stable diffusion on a local machine and the end goal is to have it available in discord.

both of those things really.

[nova zodiac]

why discord specifically?

[marble steppe]

because I'm already here messing with midjourney and I kinda like it tbh.

[nova zodiac]

in that case your steps will be to find a tutorial on youtube for setting up automatic1111, then installing https://github.com/AndBobsYourUncle/stable-diffusion-discord-bot

you'll get way more options using the web interface rather than discord, and there are plenty of tutorials on how to use it

but there's enough info there to point you in the right direction

stick with ubuntu to start with, as it's most common and probably best supported

[grizzled mortar]

hey @nova zodiac have you installed sd unclip before on automatic1111? Curious to see if anyone has done it with any luck

[marble steppe]

someone before mentioned a different distro but I can't remember the name....might've began with a z?

[forest trout]

I'd argue that this one is better:
https://github.com/Kilvoctu/aiyabot

[nova zodiac]

not personally

[nova zodiac]

Never used em, happy to hear there are recommended ones out there, I just took the first one I saw on google 😛

Honestly, at this point I should start a "SD Tech Support with ResidentChiefNZ" service and start charging people 🤪

[marble steppe]

I

m mainly just exploring options right now

I'd like to use the webui and discord if possible. for the reasons you stated that the webui has more options.

Arch Manjero, thaat's what someone said previously.

[scenic thunder]

Why don't I have permission to speak in the bot of stable diffusion?

[nova zodiac]

Because #1047610792226340935 - they are off for now

[scenic thunder]

[hallow sable]

does Dreambooth only work with 1.5 models????????????????????????????????

been trying to get it to work with 2.1 768 res models and it generates sample images but when I try to load the checkpoints to says runtime error and wall of text about nearly every res not working... it only loads if i copy the 768 ylam txt file and rename it to the safe tensor model ive generated in dreambooth!

[vestal dew]

anyone know a SD app/extension to make those talking head, animated head video? Is there something else that can do that?

[grand radish]

Does anyone know why VAE's won't work for me in SD webui?
I have downloaded all the top common/recommended VAEs and selected them manually.
(command line i am using is --xformers --no-half-vae)

sd webui throws no errors either:

Applying attention optimization: xformers... done.```

*A key point is the image doesn't change at all when I select VAE: none, anythingKlF8Anime2VaeFtMse840000_klF8Anime2 or vae-ft-mse-840000-ema-pruned or any other of my VAEs

What might cause this behavior?

My images are blurry and greyed out and terrible...

[nova zodiac]

SadTalker

[vestal dew]

thank you

[nova zodiac]

No, but noone is doing anything with sd2.x models any more, either 1.5 or sdxl

[nova zodiac]

In the settings- are you using Automatic or a named VAE? And are you loading a 1.5 vae on a sdxl model??

[hallow sable]

does SDXL work with dream booth???

[nova zodiac]

It should, if not use kohya_ss or onetrainer

[hallow sable]

kohya ss or onetrainer is not local is it? Ive heard some people using them but all the videos are done with server

[nova zodiac]

They can both be locally installed or run on runpod/colab

[hallow sable]

ok thanks very much I will try to DL and run SDXL with dreambooth, also look into kohya ss / one trainer. its been very hard to find info I've been spending the past 3 days trying to figure this out, most of the info is outdated or not all there

[nova zodiac]

Id recommend starting with kohya, thats the one likely to have the most tutorials on it

[hallow sable]

can I use filenames like how Dreambooth works ?

[nova zodiac]

I believe so

There is a #🔧｜finetune which may have more info for you if you get stuck

[still shoal]

Hey, why I don't have permission to use the bots to create image?

[nova zodiac]

#1047610792226340935