#💬｜general-chat

1 messages · Page 54 of 1

vast ingot
#

it's a hard question, because I'm not sur what you mean with "what version of stable diffusion".
Stable diffusion is the inner brick inside automatic. it has not a lot of versions.
The models are what we talk the most about, version wise. It came out on 1.2, and is currently on 2.1. But this is where you choose your model in the UI, top left.

#

damn, you deleted when I posted

#

I hope you still get my response x)

vernal otter
#

But thanks!

vast ingot
#

no flaming going around, not under my watch RememberTheTOS :p

#

but happy you found it though and thanks

vernal otter
vast ingot
#

people doing the effort to look for themselves can stay quite rare you know

#

so ❤️

vernal otter
#

Here's one I've been chewing on the the last hour. I downloaded a model called CarHelper.pt, dropped in the folder that says "put stable diffusion checkpoints here" and CarHelper.pt didin't get recognized after restarting AUTOMATIC1111. So I thought I'd be smart or not and modify the file extension to CarHelper.ckpt and it appeared to load ok, but then errored on my first render.

normal pasture
#

I am using the replicate inpainting api to generate product background for my spice box and other product (bottle of lotion for example). Unfortunately the AI keeps sticking objects to the spice box even with a perfect mask. I would love hearing from somebody if there is a way to fix this problem 🙂

brisk bloom
#

no renaming

kind kraken
#

My 4090 is arrive Tuesday 🤞 along with upgraded, well everything.

#

What are the few models that I should try running on this beast?

proud locust
#

So I have a project that I would love to mess with. I have a bunch of old-school Pin-Up artwork that was drawn by original artists from the WWII era. I would love to train a model using them, but I would preferr that it is not the original art itself that comes back when it is prompted. I actually would love to see what it looks like when I try it on top of different models, in hopes that it would keep it's general style. Knowing that each peice of art is not of the same person, some have backgrounds, others do not, should I be doing a textual inversion or a LORA? Should I train for style?

vernal otter
severe gazelle
#

hii, did you know the developer discord username?

crisp verge
#

nice

hidden fog
severe gazelle
hidden fog
#

np!

severe gazelle
# hidden fog np!

casually you know if he is in any discord server or community in where i can be?

hidden fog
severe gazelle
#

why not?

hidden fog
#

i just dont idk

kind kraken
#

Is there a discord group for control net?

finite bobcat
#

i tried img2img and i can't get my image to look close at all to original. say i have a photo of myself or someone and i want to turn them into a cartoon. i drag an image in and prompt i put something like. Cartoon, 90s, anime. i've been getting pretty bad results. any suggestions to improve my results

maiden tapir
#

Gmgm

#

How are you doing today fam

willow coyote
vestal dew
#

Today I explained to a student how to understand the main idea of paragraphs in a passage. I realized was was describing a process like LORA training to him - look at each paragraph, remove what's common, what's left is what's unique, the main idea.

stuck furnace
#

will sd ever be better than midjourney? right now it sucks shit

alpine bear
hallow plaza
#

Is the an SD model that will give me outputs like midjourney?

stuck furnace
stuck furnace
fathom willow
#

Hello guys

alpine bear
# stuck furnace ahaha nah its just dogshit

Have you not seen the things we are making with sd? It is far more customizable than midjourney. If you're using the base SD model and webui, then you're severely limiting yourself. And if you haven't even tried it, then your judgement carries no weight merupog

A few ways you can get the most out of SD:

  • Install a custom webui like Comfy or Auto1111
  • Use custom models available on CivitAI and Huggingface
  • Explore Textual Inversions / Embeddings
  • Explore LoRA adaptations
  • Explore ControlNet & Latent Couple Extensions
  • Train your own models/embeds/loras to replicate a style/subject of your own choosing

There are a plethora of phenomenal resources available in this server to get you started. Coming in here and saying that SD is dogshit... well, it says a lot about your integrity. It is nice to be anonymous.

alpine bear
dapper forge
#

Morning

stuck furnace
#

Aint reading allat

alpine bear
stuck furnace
#

💀

vestal dew
#

I have explored Lora

#

cant get it installed to make them

alpine bear
vestal dew
#

local

#

koda or khola or ehatever it is

alpine bear
# vestal dew koda or khola or ehatever it is

kohya, gooootit. There are two variants out there - one is specifically for google collab. I followed Aitrepreneur's video on training lora, was able to get it working by following his steps precisely.

calm mist
#

Does this server provide any bot for creating an image???

alpine bear
vestal dew
calm mist
alpine bear
alpine bear
vestal dew
#

the link, and then he does something with setting permission for the cmd prompt

#

er, powershell

alpine bear
#

were you able to get the webui up and running?

#

guiz is online, he gon be mad I'm doing tech support in gen chat owoShy

vestal dew
#

I can't find the powershell script he used, so didn't get past that

#

I well go to the right channel

alpine bear
normal pasture
#

I am using the replicate inpainting api to generate a product background for my spice box and other products (bottle of lotion for example). Unfortunately the AI keeps sticking objects to the spice box even with a perfect mask. I would love hearing from somebody if there is a way to fix this problem 🙂

alpine bear
lost swift
#

I have small question to ask

vast ingot
#

ask away 🙂

lost swift
#

since I change the weight of this art style in xyz plot.why there is no difference?

vast ingot
fathom willow
#

How to generate

surreal perchBOT
#

@fathom willow

FAQ: How do I generate images? Is there a bot on the server?

Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝｜tech-support for any problems you encounter while installing it locally!

lost swift
vast ingot
# lost swift Thank you, I have update it

and it's a text inversion ? a lora ? or just the base model ?
In each and every case, it's really strange, it's as if there was no change at all in prompt submited to the model there

#

if you drop the individual pictures from your output folder into PNG info tab, it will show the prompt that was ran for that picture. can you confirm it used different weights for real ?

#

Trexdel gave you some good tips too

lost swift
#

I see

#

Thank you very much

normal pasture
#

@alpine bear Hi Trexdel,

Thanks for sharing your approach to dealing with the issue of objects sticking to product images 🙂

I was curious if you had any knowledge on how Pebblely.com is addressing this issue with their AI-generated product images? I was wondering if they are using any specific techniques or methods to prevent objects from sticking to the product images...any clue?

alpine bear
# normal pasture <@243730375845740545> Hi Trexdel, Thanks for sharing your approach to dealing w...

I'm guessing their models were trained on 'empty' images. A dataset of images that all have empty space in the middle for product placement could easily replicate the effect.

The best thing you could do to achieve this in SD... well there's probably a couple things. A custom LoRA with extremely diverse backgrounds, all with consistently empty space might work. Controlnet might have a use case here as well. I haven't tinkered with this specific use case before, so I can't speak to the efficacy of my ideas sfomegalulcat

normal pasture
#

@alpine bear Good gues! Thank you very much!

native sandal
#

😃

zealous narwhal
#

I have a question for locally run Stable Diffusion

#

If i want to install this https://huggingface.co/runwayml/stable-diffusion-v1-5/tree/main

Normally i just dl the safetensor or cptk and copy it to my models folder, but what are all these other folders and Do i need them and how and where do i install all these dependencies? (Like from a GitBash Terminal preferably)?

latent thunder
#

One question, do I have the copyright to market the images that I create with dream studio?

hushed quarry
wintry echo
#

Hi All! Im super excited to be here. Building my leadership and advisory council for Assemble Teams Inc., enterprise AI software. Love to chat and connect sometime. Based in California

latent thunder
hushed quarry
hot marlin
latent thunder
fiery pecan
#

what does CFG Scale mean ing image2image?

vast ingot
near silo
#

Alright, so, if I were to make a guide for how to replicate my high quality results out of ultimate upscale. Where would you all recommend I upload it to? I am not sure how to go about this, but this process works so go that I wanna share it

willow coyote
near silo
fervent thunder
#

i wonder is stable working on any kind of chat bot. i assume they must be

willow coyote
#

Oh you're writing a guide?

#

Wait until you see the images it creates. It's for what upcoming world's most popular record label 🤞

#

They'll be better than anything ever posted in here.

fervent thunder
#

which model you using

#

or version rather

normal pasture
fervent thunder
#

where does the safetensor file go

warm junco
normal pasture
#

I am really desperate...may anybody help me?

inner iron
#

?

normal pasture
inner iron
#

have the same problem^^

#

but if u have an excat mask it will do it right i guess

normal pasture
#

Are you also using the replicate API?

inner iron
#

but masking is nat so a big deal i nSD

normal pasture
#

No, unfortunatly it will still attach objects to my upload image...

leaden mural
#

How do you do control net with hands?

normal pasture
#

Am i using the wrong color code for the mask, using:// Set the white color for the mask
$white = imagecolorallocate($mask, 255, 255, 255);

// Set the black color for the mask
$black = imagecolorallocate($mask, 0, 0, 0);

devout hull
#

what port is the backend of sd using?

fervent thunder
#

is there a command I can use to download these massive models from civit ai?

#

My current method is DL to my desktop (4-5 hours), upload to google drive (4 more hours), and then use the gdown command to download it from there

#

!w get doesnt work either because it will download the file, but without the filetype. Just a blank file. No ".ckpt" or whatever

worn hound
#

Anyone know what happened to unstable diffusion?

#

After that drama

fervent thunder
#

Please help I have to sit at my computer for like 5 hours straight just to get one model working

long sigil
#

how can I generate a logo with specific letter ?

topaz fern
#

im tryna make friends, pls someone add me and be my friend

normal pasture
#

Anybody understanding how Pebblely is working?

#

This is a really greate opportunity to make money

inner iron
long sigil
warm junco
long sigil
fervent thunder
#

I did try that in the past

#

And I added the signifier at the beginning of the prompt like this

#

"1990s (style)"

#

and it didn't seem to effect anything

#

any clues?

long sigil
#

BRO

#

img2img added a watermark lol

#

to logo

warm junco
cursive gate
#

Is anyone using control net with 2.1?

drowsy verge
#

Flying lion#

distant quail
#

oki

jagged solstice
#

Hey, im having trouble getting stable diffusion working locally on my new amd build...could someone give me a hand?

thick gorge
#

Is there a way to rent VM online with windows 10, to run Automatic1111 there?

dusk jasper
#

i need help to create art someone could f give a hand? with stabble diffusion

wintry stream
jagged solstice
#

cheers

broken smelt
#

i keep seeing posts about an chatgpt extension for prompting, is that even effective at all? good prompts usually have a formula to them where you dont want to type "draw a beautiful woman holding a phone walking by the beach and her hair is long and blonde" but instead you want to be specific with i.e.: woman holding phone, beach scene, long blonde hair, etc. Can't imagine chatgpt knows how to type that out every time. seems faster to just parse it out yourself instead of typing a non-working prompt to the plugin lol

mortal zealot
#

Feed documentation into the gpt4 model and then define the formula you’d like for it to use every time.

fervent thunder
#

are there any open music/sound ai communities yet?

honest blade
hallow cypress
#

Yes. It is!

dense tinsel
#

Hello

hallow cypress
#

It seems that problem started around 5:30 UTC

dense tinsel
#

Is there any stable diffusion extensions on chromium?

honest blade
#

Ah, thanks. Is there somewhere SD posts updates on this?

hallow cypress
#

I would like to know too!

#

Also, Stability's API page does not seem to work either. Just me?

visual musk
#

Finally figured out that it's not my code; it's the API

#

What are some other good Stable Diffusion hosts? Are there any that are competitive with the Stable Diffusion API? Replicate? Banana? etc.

hallow cypress
#

Great, the API is working again

hallow cypress
visual musk
wild steppe
eternal geode
#

How different is sdxl from sd 2.x?

#

By 2.x i mean 2.0 and 2.1

hallow cypress
honest blade
#

It's still not working for me...

hallow cypress
#

It is very slow. But it is working. 20s or more per inference

visual musk
#

It's been on-and-off for hours, so I wouldn't expect it to be consistently working until staff like Jae says it is

hallow cypress
#

We should tag Jae so that he can tell us that the API is working

karmic brook
#

there will be an update as soon as there can be and you all will be informed of that--no need for tags ^^

hallow cypress
#

Thank you

visual musk
#

@hallow cypress @honest blade While we have some Stability API users here, what are y'all using for async calls to the API? I had to stop using the python stability-sdk and, instead, write my own that uses aiohttp with the REST API in order for my application (discord bot) to work.

hallow cypress
#

C#'s IEnumerator. Earlier I used to use the main loop for the call to the API.

visual musk
#

Ah, maybe it is better in C#; nice. Luis, are you using Python?

honest blade
hallow cypress
visual musk
#

Maybe it will becoming it's own GitHub repo if others want it 🤷; doesn't seem like that many people are impacted by the API outage, though

hallow cypress
visual musk
astral goblet
# normal pasture regarding this:I am using https://replicate.com/stability-ai/stable-diffusion-in...

this is a toughy. Since you want the background to know the masked part is a wine bottle, you should prompt "wine bottle against background" sort of prompts, which i assume you are. Since it's adding a lid on top of your wine bottle lid, or making the bottle look wider. I'm not sure if this would work but maybe, "wine bottle" in the negative prompt field would help?

If we were using the webui made by a1111 i would suggest you try to use prompt editing syntax for some nifty tricks. [wine bottle::0.2] in prompt and [:wine bottle:0.2] in the neg. I have no idea if that'd help but i would try it as an experiment. It would only have wine bottle in the prompt for 20% of steps, and then put wine bottle in the negative after 20% of steps.

maybe other people have tricks for this? you'd want the inpainting to know the context of the subject, but also not extend the subject

fervent thunder
wind heron
#

Hey guys, i am trying to install / use Kohya_ss, The video from Aitrepreneur is outdated and the intrustions are no longer the same on the github, so im a bit lost.

normal pasture
#

@astral goblet thanks for your detailed response 😉 I tried it with your suggestions…unfortunately it did not help at all :(. Does anybody have a clue how make this work properly?

burnt lichen
normal pasture
#

#everybody I would be willing to pay if somebody finds a good working solution$

burnt lichen
#

I'm pretty sure the endgame for this problem involves 3D, and probably a completely new approach to building the system from the ground up. Blender + ControlNet is a step in the right direction, but I think you can't get to there from here without considering UV coordinates or Generated coordinates or some sort of 3D data, instead of just screen pixels of the output image.

fervent thunder
burnt lichen
#

I've tried clearing my cache and turning on a VPN. Neither fixes the problem.

pulsar lantern
#

hey guys im facing this issue on google colab.

RuntimeError: Detected that PyTorch and torchvision were compiled with different CUDA versions. PyTorch has CUDA Version=11.7 and torchvision has CUDA Version=11.8. Please reinstall the torchvision that matches your PyTorch install.

anybody got a solution?

astral goblet
quasi drift
visual musk
tribal stream
#

Ello peeps I am back I lost access to my @cedar garden account by being a idiot lmao 😂

willow coyote
queen spruce
#

Anyone have experience with an Nvidia HGX A100 80GB 4GPU Card? Are these types of cards picked up and used by Stable Diffusion?

#

vs like the normal PCIe A100, for instance?

brisk bloom
queen spruce
#

I know, this is in the context of Lenovo server

#

Assuming I have this hardware, I should be able to install SB on it and have it picked up just like any other GPU, right?

brisk bloom
#

no idea, sorry

errant solstice
astral goblet
#

Yeah. Thats the kind of juice SD loves to work on

#

its a custom nvlink enclosure for 4 a100s if i'm not mistaken. i think its the kind of hardware that base models are trained on

#

yeah stability has a 4000 a100 cluster back when i read first about them

visual musk
random pumice
#

Yall able to go on civitai?

#

Website looks like it's having trouble

vestal dew
vestal dew
latent thunder
#

Hello there, when I use Dream Studio if I generate image via API, and the Finish-Reason is CONTENT_FILTERED do they charge me the credits?

untold ocean
#

no, credits are not deducted for generations that are flagged by the content filter

burnt lichen
#

Does anybody here know how to use Charturner with a normal map baked from an external program? I have Charturner, I have my prompt, but when I try to use my prebaked normal map with txt2img, it spits out two characters even though the control image normal map clearly shows 4 characters.

hexed hatch
#

ok i have a extremly quick question for anyone with auto1111
its not even really technical or complex, just anyone with it installed
do you guys know which folder "Lib" is supposed to go in?
when i was freeing up space, i removed this lib folder separately

burnt lichen
hexed hatch
#

specifically the lib folder with like hundreds of thousands of files, ill get a ss quick

burnt lichen
#

Oh wait. They're all in \venv\

hexed hatch
burnt lichen
#

Yeah I don't have nearly that many, but \venv\ seems like a safe bet for the parent folder of \lib.\ So \venv\lib.

#

Also I think if you're using a physical mouse you can click the path and drag to the right and see what's at the end of that string.

hexed hatch
#

ok thank u for the help, ill see if it works

burnt lichen
#

C:\Users(your user name)\stable-diffusion-webui\venv\Lib\ if you did a default install.

#

replace (your user name) with the name you type to log into windows.

hexed hatch
#

alr i think what u gave works, its just some other parts of the program not working

#

so ill just try to reinstall
edit: i got it working, all that i needed to do was add the empty folder specified above,
since i deleted empty folders recently it didnt work

burnt lichen
#

Good luck.

latent thunder
queen spruce
#

@vestal dew it was $130K actually. Nvidia said they'd get an answer and it's been 2 days

latent thunder
#

Hello, I read the prompt token length of dream studio is 75 tokens.
How do I know how much is 1 token? Is there any algorithm or public library?
I have an app that currently uses the dall-e api and I will migrate it to the dream studio api, so I want it to throw an error if the user prompt is longer than allowed before sending the request to the api.
Thanks

vestal dew
queen spruce
#

it seems like such an easy question to answer... my guess is... we didn't spend enough money and no one cares

vestal dew
#

have you called multiple times?

astral goblet
#

youknow those money boxes where money's jsut flying around and you grab as much as you can? i think that's what nvidia sales departments are like lately

alpine bear
#

That sounds sketchy af.

astral goblet
#

ew.gif

alpine bear
#

Oh, oh no I see the pfp

#

TOS

hoary river
#

Is the webui fixed for sd?

fervent thunder
#

So are prompts heavily checkpoint reliant

#

Or are there some good FAQs to get better at entering baseline prompts

#

Been playing for two weeks and have a basic understanding now and it's time to study and get better.

warm junco
fervent thunder
#

How does () work?

#

If I put my (word) in like this does it make the prompt more virile in the processing

#

Also how much of a difference is front of the prompt lines compared to back. Front,Middle front,Middle back,back

#

Is there a specific percentage difference in severity of output from front to back?

#

Also what percentage it decreases depending on the amount of prompts?

brisk bloom
fervent thunder
#

Ty

quasi drift
burnt lichen
#

Actually, I'm seeing a lot of random projects on Reddit popping up exploring the integration of 3D. Blender. Unreal Engine. Pretty much anything with an API.

quaint trellis
#

Hi, I just tried installing the application and encountered an error. I don't have an NVIDIA GPU; I only have an AMD GPU. Here's the error message I received:

Traceback (most recent call last):
File "C:\Users\Joshu\OneDrive\Desktop\School\Art\AI Art\stable-diffusion-webui-master\webui.py", line 139, in initialize
modules.sd_models.load_model()
...
RuntimeError: Found no NVIDIA driver on your system. Please check that you have an NVIDIA GPU and installed a driver from http://www.nvidia.com/Download/index.aspx
Can anyone help me understand how to resolve this issue or provide guidance on running the application without an NVIDIA GPU? Any help would be greatly appreciated!

warm junco
#

You have installed the wrong version of the webui

quaint trellis
tawdry turtle
tough panther
#

Hey! Do you know of a discord for Lora Training?

lavish current
#

Why are there so many people on this server who hate AI

abstract flame
dapper forge
#

AMD support yet ?

shadow birch
#

is the poseX extension bugged by google colab? when I open the tab, only a checkbox appears to send to control net

fervent thunder
sage shale
hallow plaza
#

If you want to train on a 768 model, can your training images be 512x768 or do they need to be 768x768? Also, can you mix in 768x512?

shadow birch
sage shale
shadow birch
solid orchid
#

Is there a way to enable multiple graphics cards at the same time to accelerate the calculation of a single graph?

#

Or at least multiple cards are called simultaneously in the Web UI to generate pictures

shadow birch
#

guys, does anyone here try the depth library to create the hands? does anyone know how to solve the problem of hands appearing dark or a different color than the rest of the skin?

normal pasture
#

is there a api which allows me to use AUTOMATIC1111 inpainting?

#

Just found img to img.

long sigil
#

why can't this thing draw a simple letter?

#

bro I just want a "letter" nothing more

#

it keeps drawing something like dragons, animals, anything that what is random

fervent thunder
#

Hi all. Im new to SD and was wondering if there is a help section to post something im stuck trying to figure out. My Loras are not loading in for some reason and my GPY isnt dumping files it has cached from old renders until I close SD and start over.

#

" sryGPU"

shadow birch
#

guys i love dreamshaper

proud nova
long sigil
astral goblet
shadow birch
astral goblet
#

it sure is boob thirsty though. better than other blends that hallucinate orifices all over people's anatomy

karmic brook
floral umbra
astral goblet
# floral umbra Time to windowshop for my next 16-18GB ram phone with snapdragon lol or exynos. ...

My advice, wait a year for a generation 2 qualcomm ML chip. When qualcomm releases new tech on their silicon, it's always SOOO beta. They have a very "tick tock" style release pattern, where the tick is something new and the tock ties it all together to suck less. When their ML features show up in phones that aren't basically foldable phone gimmicks, and are in models that consumers rather than enthusiasts can afford, you're probably good to go then

floral umbra
#

Indeed. Time to keep an eye out!

#

I do need to replace my phone's battery as planned obsolescence got my S21 on near crutches atm

astral goblet
#

lol QC has even got wise to purchasing stratagies i'm suggesting. The chip they got SD working on first is called the Snapdragon 8 Gen 2 Mobile Platform LOL. They know people are going to wait for their gen2 devices to buy them so they just call the prototype gen 2 lololol

shadow birch
#

how do you make the depth library work right? it keeps showing the character holding an object instead of showing the hand in the right position of the model

astral goblet
#

i got a Snapdragon 778G right now. It's a "5g mobile platform" lol but it doesn't do 5g networking HAHA. the marketing brand names out of qualcomm are so dumb sometimes

tawdry turtle
#

Why can’t I generate nsfw pictures through DreamStudio or api?

amber jay
#

I can't prompt anywhere

#

Why are all rooms locked ?

near silo
#

@amber jayThere is no way to generate in this server

surreal perchBOT
#
FAQ: How do I generate images? Is there a bot on the server?

Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝｜tech-support for any problems you encounter while installing it locally!

astral goblet
wheat anchor
#

I am writing an article about AI image generation for my technical writing class

#

I wanted to ask if there was any good resources about the history of AI image generation and how AI image generation works

vestal dew
#

It began at least 5000 years ago

#

the first well known image generation computer program was called Aaron - 1973

topaz blade
#

I am looking for a mentor, that helps me with Stable Diffusion. This thing is really overwhelming, and even though I manage to do some prompts, I suspect I am just scratching the surface...

still mesa
#

ya'll how can I donate to stability ai so they can rekt midjourney for ending the free trial

#

?

fervent thunder
#

Hi guys, sorry if this is off topic, but maybe someone knows Ai tools to make a 3D model of the body and face from a photo?

shut lantern
#

can someone set up my google colab for me thru teamviewer plss like move everything i have from local into google i can pay

latent hawk
#

Was in the news recently:

humanity is faced with a choice that surpasses in importance the consequences of all the previous historical forks:
A. Accelerate the development of AI based on large language models in order to use its rapidly growing intellectual power to solve the most important tasks of mankind that are impossible for humans.
 B. low down and deal with the risks of further growth in the power of AI so as not to run into irreversible catastrophic consequences.
There is no polyphony of voices in China, and there cannot be. It is for the Chinese Communist Party to decide and for everyone else to accept. And the day before yesterday, the Xinhua decision was officially published, strongly resembling China’s final choice: option A.

Any opinions? Can non-democratic China actually achieve what neo-luddites in the West are trying to veto?

halcyon python
#

Hi there
could someone explain me with few words the differences between:

  • model
  • lora
  • checkpoint
  • hypernetwork
latent hawk
kind kraken
#

Help!
Followed the steps for stable diffusion installation in windows , double@clicking on webui-user.bat, just open and closes the command window very quickly

vestal dew
#

open a cmd window in your stablediffusion dir

#

in the location line of the window, type cmd

#

then in the cmd window that opens, do "webui-user.bat"

#

it will ruin the bat in the already open cmd window, which then hopefully will stay open if it doesn't run, so you can see the error

kind kraken
#

I had to redo all the steps, now it works 😀 earlier I was using Wsl now I am not, seems that was the issue

vestal dew
#

wsl?

astral goblet
#

windows subsystems for linux

vestal dew
#

I see

maiden crystal
#

hi

#

just stopping in to say hi. and now, goodnight.

broken smelt
#

Infinite prompt length
Typing past standard 75 tokens that Stable Diffusion usually accepts increases prompt size limit from 75 to 150. Typing past that increases prompt size further. This is done by breaking the prompt into chunks of 75 tokens, processing each independently using CLIP's Transformers neural network, and then concatenating the result before feeding into the next component of stable diffusion, the Unet.
For example, a prompt with 120 tokens would be separated into two chunks: first with 75 tokens, second with 45. Both would be padded to 75 tokens and extended with start/end tokens to 77. After passing those two chunks though CLIP, we'll have two tensors with shape of (1, 77, 768). Concatenating those results in (1, 154, 768) tensor that is then passed to Unet without issue.
man this is kinda tough to read. does this mean that anything past 75 isn't actually fully utilized? not sure if it makes a difference if i should be using like 110 tokens for example compared to just 75

odd geode
#

morning

fervent thunder
#

how do I generate depth maps for stable difussion?

modern ferry
#

hello

#

does stable diffusion has a logo?

warm junco
hard sage
#

new here, idk if this is the right place to ask... is SDXL beta available via API yet?

modern ferry
#

hello, kinda new to the stable
I downloaded some poses but its a zip file and don't know which file would i put, can anyone help please?

magic citrus
warm junco
modern ferry
#

how can i add the extension? @warm junco

warm junco
modern ferry
warm junco
modern ferry
dapper forge
#

Morning

echo nymph
#

Hello all how is everyone

vast ingot
#

Hello and welcome around !

#

quite good, big sunshine in my face to wake me up; love it

#

how are you ?

echo nymph
#

im apparently using an old version of SD lol but doing fine

astral goblet
#

peopel are madly training alpaca models with lora to create chat assistants

tender hatch
#

Not sure if this is the right place to ask, but does SD have an option for converting images into "pretty" alternatives? I specifically want to turn a picture of a church into a scary halloween image for a website, but not sure where / how I can do it.

astral goblet
#

this stuff is so fire to watch

#

i may be able to get one working on my machine

marsh rock
#

So when I upload image to dream studio - write the prompt and hit dream. It just makes 4 versions of my photo and does not alter it.

obsidian current
#

When making a Lora, the trigger is just the file name yeah? lora:ExampleStyle:1 so in the prompt I'd need to type 'ExampleStyle' or am I wrong? Using Koyah ss. This is my first try at Loras and I can't seem to get it to do the style right that I'm training on without a trigger word.

proud locust
#

I need some quick advice. I have a bunch of tshirt and sticker designs that I would love to train into a model in hopes of allowing the ai to use what it sees and then reinterpret it into new inspirations. Basically, I don't want the prompt to just pull up the exact graphic that it was trained with. Is that possible by training a textual inversion?

jaunty light
#

sadly there is a 3rd leg testing new models

fallow swallow
spark phoenix
#

Hi guys, when i try to generate an immage a error message appears: "OutOfMemoryError: CUDA out of memory. Tried to allocate 64.00 MiB (GPU 0; 4.00 GiB total capacity; 2.63 GiB already allocated; 40.98 MiB free; 2.69 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF". How can i fix it?

fervent thunder
#

hii guys pls i need help

#

when i try install stable

#

i have this error: ''exit code: 1 ''

#

how do I solve it?

warm junco
warm junco
astral goblet
#

tiled vae is a new extension that i'm quite impressed with and i'm wondering if low vram users can benefit from it

#

generated a 1440p wallpaper and it topped out at 8gb of my 16gb, which is extraordinary. I usually max that on those gens and often error out

warm junco
#

Ah wow would be nice to have a Tutorial for that

#

Looked at it but didnt tested around much

astral goblet
#

i just enabled it and went for it with default settings

warm junco
#

Does it change the colors ?

astral goblet
#

not really no. i used it with illuminati diffusion 2.1 and it worked great

warm junco
#

Okay good to know

astral goblet
#

i didn't really test seed vs seed with and without though

#

oh neat! discord doesn't just use the big raw link now. it ties it into a tidy little channel name link

warm junco
#

Gonna test Tiled vae later again hope it helps with my 8gb

astral goblet
#

in the past i wasn't able to easily make 2k wallpapers. i had to a hires fix with low denoise and hope my gpu was feeling efficient and it would complete the pass after many minutes. sometimes it would get oom error and continue anyways. Struggled process at the limits of my card.

with tiled vae i can just generate them without a hires fix! but that doesn't work well still since the 512x focus problem. Even so, now it generates them without it! and even when using it, it's faster than before! likely because the memory swapping happening with the tiled vae is so much more efficient than what happens at the limits of my card under normal operation

#

oo my bad. dropped last night

#

okayokay hollddd up. the author be like "It's not based on stable diffusion" and is saying this is a base model

warm junco
astral goblet
warm junco
#

Hmm okay

warm junco
astral goblet
#

oh my bad. the post was just from the dreamlike.art owner. The model was created by another team i think

rough barn
#

is there an advantage to hires fix over just using the SD-Upscale script?

astral goblet
rough barn
knotty turtle
#

Anybody tried "Seams fix" in Ultimate SD upscale with large resolutions yet?

My most recent upscale took 1hr 18m and I didn't wanna re-run it with Seams fix on if it doesn't work well. But man, I need it. I got seams bad.

astral goblet
#

i wish though that we could do multiple seam passes as sometimes they don't blend as well as i'd hope. lower denoise helps

#

ultimate upscale is fun to toy with. it's built really well for a first generation diffusion upscaler

#

maybe 2nd gen?

knotty turtle
astral goblet
#

you don't have to. you can put any prompt you want. The thing is though, the actual diffusion is only happening on tiles of the larger image at a time. So if your prompt is about the main subject, it'll try to put the subject on every single tile

knotty turtle
astral goblet
#

i like to use prompts about detail instead. if its a photograph i talk about photographic details and terms. if its a painting i talk about brush strokes and textures.

knotty turtle
astral goblet
#

should just be a low denoising pass too. i ride around .25 and .3 getting out of that tends to hallucinate too many new details

knotty turtle
#

I've been running canny, scribble and depth on my illustration upscales, but i can't quite tell if it's helping

astral goblet
#

no. i don't know how to integrate that really, since the script splits the image into many tiles and works each pass on a tile. controlnet would have to integrate in somehow and split it's control image over the same tiles, but then you might get a lot more artifacts

#

What i have done is take a final hires upscale, select and cut out a tile in an editor, rework that in the ui with a controlnet image of the original matched against it. inpainting better finger tips and other details. then taking a few good results back into the photo editor to stitch them back into the final piece. i've done this process twice with diffusion and i could probably improve it, but that sort of janky manual work i'll do a lot of

#

it's hard to get stuff to line up sometimes with controlnet

#

and with the tools as they are, it's hard to see if they're not entirely lined up

knotty turtle
fervent thunder
#

whenever WebUi finishes downloading it just says, "DiffusionWrapper has 859.52 M params." and wont give me a link. how do i fix it?

#

so what should i do

#

how long should it take?

warm junco
fervent thunder
#

can i send u a screenshot in dms

#

of my cmd bar thingy

warm junco
#

Yes but today 4 people got stuck at that point

#

Something is wrong

#

One guy waited 3 hours

#

It was a fresh install for all of them

#

Python path was an error for one of them

#

Yea but it didnt happend before that much at the same time

#

Yea it isnt an error

#

Its not

#

Its a big problem today as you can see here.
Its an issue for a lot and not a common thing

sand pawn
#

Hey everyone, is there any roadmap page for upcoming stability.ai SD models? (the latest available being unclip_2.1)

astral goblet
#

SDXL is next to emerge. larger param set. a russian company released their image diffusion system today too. large param set there

sand pawn
#

thanks

still glacier
#

"Escalation of commitment" is another bias.
We've been having multiple people for the past day or two coming and saying they've waited hours. fixing their etag just like mentioned in the github ticket fixed the issue.
So either they all decided to lie at the same time, or there is indeed a problem.
For sure Auto1111's UI design is not the best and tons of fluffy loading bars could be added to let the user know what's going on under the hood. But Auto1111's repository has never be focused on "ease of use".

#

hijacked github ticket maybe, probably

astral goblet
still glacier
#

thanks have a nice day

astral goblet
spark phoenix
hazy wind
#

Will the millions of images removed by artists significantly affect the quality and fidelity of the new SD version?

shadow birch
#

Guys, does anyone know of a tutorial to install those packs of negative prompts that are placed in the stable diffusion folder, and then just use keywords to invoke them at the prompt?

silver osprey
#

question,is there any way to use stabble diffusion on mobile?

shadow birch
silver osprey
#

how?

shadow birch
# silver osprey how?

the same way you use it on a pc, just access the colab links in your cell phone browser, and do the same installation process

shadow birch
#

there may be another way, but then I don't know

silver osprey
#

yeaa do you have a tutorial or anything?first time trying smth like this

shadow birch
#

very simple brother, open your mobile browser, access the git hub camenduru link, and choose the SD version you want from the list, then everything is intuitive. you just need to login with your gmail account on google colab website

silver osprey
#

what camenduru link?

primal marten
#

Hi there! Perhaps this is a silly question but what's the difference between installing SD locally vs installing it with Google Collab?

#

Is one more convenient than the other?

shadow birch
silver osprey
#

yea im so fucking tired i dont understande shit here

hazy wind
#

So the answer to my question is "No"?

astral goblet
#

why even ask if you didn't want a good answer? ugh. people annoy me today. it's just so senseless to be that rude about writing that much for you

potent spire
#

I dont think the quality will drop massively because of that

#

Adobe Firefly does a impressive job with a radically smaller database than SD or MJ

#

So the number is not all

sage musk
#

is it better to use a higher resolution or higher sampling steps?

astral goblet
potent spire
#

Adobe used only Adobe Stock and license-free/ license expired images

#

Midjourney basically scraped what they came across to just like Stability AI/Stable Diffusion

astral goblet
#

Which can be quite a lot. Adobe has been in the phtoography game for quite some time. Adobe stock also includes all images hosted on bridge

potent spire
#

Yeah, but they dont use anything outside of their area and what is "ethical"

#

S-AI and Midjourney did scrape what they came across

astral goblet
#

my point is, we can't confidently make claims about firefly OR mj's dataset. they're closed sets.

and i hate when people put quotes aroudn ethical. I can tell you have disdain for the subject now. ugh.

potent spire
#

So ofc they will have a larger dataset

astral goblet
#

people annoying me so much today. i'm checking out of this convo too. quotes on ethical. UGH

potent spire
#

Well this is the third (?) time now you are annoyed by someone and you still answer to all 3 ^^

astral goblet
#

don't twist words. be less childish.

potent spire
#

Bro, you are the one having a bad day, not me

astral goblet
#

weird

warm junco
#

But it depends on what you want to generate

warm junco
sage musk
#

So i think staying under 1k in either dimension + upscaling would be best right

warm junco
#

To get nice images

sage musk
#

Damn i hoped i could utilize our 24 gb vram xD

astral goblet
#

hires fix in a1webui tries to solve the resolution attention issue by starting with a small scale generation first, and then doing a 2nd pass on that at the desired resolution

#

it's a cheap hack but it really does provide great generations

warm junco
sage musk
#

Interesting

#

I should read about batch size, i only used batch count until now

warm junco
astral goblet
#

with 16gb i've got as high as 80 512x images at a time

#

mmmm

sage musk
#

And that's faster in total?

astral goblet
#

think it took less than 2min

warm junco
warm junco
astral goblet
#

i have but i haven't pushed those to high batches yet. for the dataset i'm generating out , i am creating 768 images that start as 512 with a hires fix of 1.5. doing batch sizes of 10 there

#

i want to play around with tiled vae and batch sizes a little too

sage musk
#

Like i can't imagine rendering 50000 images for the price of one if there'd be enough vram

warm junco
#

Thats just an example so you could do clearly more images

bleak wolf
#

Why did Emad sign against GPT 5?

sage musk
#

So that's just the question, if i theoretically had infinite vram then could i generate infinite images in batch in the time of rendering one normally? Logically there would have to be another limitation by the gpu cores

warm junco
shadow birch
#

does anyone know in which folder to put the sd-vae-ft-mse files?

warm junco
shadow birch
brisk bloom
shadow birch
#

Guys, I know you have to take care of the models we use in stable diffusion to not get hacked. But what about those who use Google Colab, are they also at risk?

stuck furnace
#

@sly escarp hey

#

I wanted to apologize

#

I did take the time to install and use stable diffusion

#

its good

#

I like it

#

I was wrong

alpine bear
#

💀

stuck furnace
#

👍

burnt lichen
#

🫰

#

(I'm frankly sort of disappointed there's no emoji of a hand with the thunb both up and down and at least five curled fingers for me to use ironically in this situation.)

brisk bloom
heavy heron
#

Hi is this the stable diffusion official group chat
I want to ask in files , can we throw ckpt and safetensors in model, or any downloaded files inside model ?

wind pecan
#

maybe not a smart question but is automatic1111 the same as stable diffusion? i already got stablle diffusion on my pc, but i cant see the safetensor files

quiet vector
#

the interrogator does not work in my stable diffusion automatic1111 install - is that normal?

alpine bear
alpine bear
quiet vector
#

when I click 'interrogate clip' the prompt box dissapears and it says in red text ERROR - the command prompt gives: {'error': 'IndexError', 'detail': '', 'body': '', 'errors': 'list index out of range'} - followed by much more info

#

ah, guess this is also tech-support food 🙂

alpine bear
wind pecan
meager shadow
#

i swear i feel like i draw all of my drawing after all the research of prompts lmao

alpine bear
wind pecan
#

if u want i can share my screen, maybe you ll know if i did a mistake somewhere

#

It looks like i need Automatic1111

alpine bear
# wind pecan It looks like i need Automatic1111

Regardless of whether you need it, I would definitely recommend it. Auto1111 webui has a much larger feature set than base sd webui. As per why your models aren't showing up, I don't have an answer for that one :V

wind pecan
alpine bear
wind pecan
#

hmm okay, thnx

merry crown
#

yooo

steel walrus
#

can anyone guide me on how to make my own models that can create backgrounds

#

consistent backgrounds of a certain quality

#

that would be great to know

foggy atlas
#

Hi guys, I'm toying with segmentation preprocessor, the goal is to get this kind of result. However when I tried out, my result, while having the segment correct, the color is really yucky. Have anyone had any issue with this?

vestal dew
#

anyone have images already prepared for a LORA or a link to such? I want to try to generate a LORA but with an image set I know works

cerulean juniper
#

where is the best place to upload my models?

vestal dew
#

civitai

torn plaza
#

why people mostly generate either women, or famous people in funny settings?

#

are people that simple and primitive

vast ingot
#

it depends who "people" is. people just starting with generative AI want to play, to have fun, and they think of lust and other small things on their mind, like making their president in stupid settings for example.

#

but once that phase passes up, if people stay in generative AI, they tend to progress towards art, or photography, in more generic settings

devout hull
#

Hello

vast ingot
#

hello 🙂

devout hull
#

is text2video even working on 8gb vram?

#

is someone using it?

vast ingot
#

not sure, there isn't txt2video in Stable diffusion as far as I know

devout hull
#

yes, but as an extension

vast ingot
#

wich one are you referencing ?

devout hull
#

Modelscope

vast ingot
#

yep

devout hull
#

I am using automatic1111

vast ingot
#

8gbs vram should be enough to run on GPU with low vram vae on at 256x256 (and we are already getting reports of people launching 192x192 videos with 4gbs of vram). 24 frames length 256x256 video definitely fits into 12gbs of NVIDIA GeForce RTX 2080 Ti. We will appreciate any help with this extension, especially pull-requests.

devout hull
#

Have cuda out of memory error... like I am missing 50mb

#

I try to lower resolution

vast ingot
#

well it takes memory in small increments, so it could be a lot more than 50mb you are missing

#

hard to know

#

but it says you should be able to run it on 256x256

devout hull
#

torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 50.00 MiB (GPU 0; 8.00 GiB total capacity; 6.85 GiB already allocated; 0 bytes free; 7.03 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
Exception occurred: CUDA out of memory. Tried to allocate 50.00 MiB (GPU 0; 8.00 GiB total capacity; 6.85 GiB already allocated; 0 bytes free; 7.03 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

#

using 256

#

it says gpu 8gb but 6.5 already allocated?

#

By what?

vast ingot
#

by the script itself

devout hull
#

ok

vast ingot
#

like I said, it allocates in small bits

#

a smaller VAE may help from what they are saying on that github (not sure what VAE is smaller though...)

#

but also, do you use --medvram or --lowvram ?

#

this could help

devout hull
#

i have no vram option

#

this is my .bat

#

--port 9091 --xformers --listen --disable-safe-unpickle --disable-nan-check

vast ingot
#

you may not need " --disable-safe-unpickle ", this removes a security measure
but add "--medvram"

devout hull
#

ok

vast ingot
#

this should reduce VRAM usage, but also slow down your image making a little

#

there is also "--lowvram" that you can use instead of --mdevram, but this is a lot slower

#

but let's move this to #🤝｜tech-support if you need more help on it, let's not spam the general channel with debug please

devout hull
#

No, fixed... thank you. Restarted pc and setting --medvram helped.

#

This is... funny

shadow birch
#

when you choose the vae in the selection field, do you need to use the trigger words in the negative prompt, or is it not accurate?

devout hull
#

Trying to merge models with supermerger, what is the best option to have most of all 3 models?

wicked timber
#

olá

dense mason
#

anyone here using both Midjourney and SD? I find Midjourney to be better at creating artistic images. Stable Diffusion is godlike at making stuffs that are forbidden in Midjourney

#

I know it costs a lot to train these models so I wonder if there'll ever be an artistic model for SD as good as V5 MJ

polar coyote
#

Anyone here every try running SD on Azure before?

potent spire
#

@dense mason used both, but MJ much more. MJ is the king atm in image quality, especially photorealism

#

Expect them to be eventually dethroned this year tho

astral goblet
slender vault
#

Being a mj user for over a year and then playing with SD I find it hard to justify the price. Especially since they just now added ()[]{} but are paywalling it behind their pro membership which would cost me 60$ a month... no thanks.

vast ingot
#

SD is less user friendly. it takes a lot more efforts to prepare a real environment in SD where you can make the same kind of art MJ does.
we get a lot of users, new ones, coming in here and ask why SD is worse than MJ.
This barrier of entry is what makes their price model sustainable too, they do quite some work to make your basic prompt into superb art

astral goblet
#

the civit community is super gross. they got a new website section for just images. it's filled to the brim with teenagers and dictators

#

can't wait for another company to step in and provide model hosting

vast ingot
#

yeah, the vocal community is quite bad there

#

but lots of people just use it to share models and loras

astral goblet
#

it's all we got for now

#

huggingface is such a bad UX that nobody will dare venture into that thicket

#

i dont know where civit gets their funding but scrolling through their new submissions , just how much thirsty junk there is. how do they pay for all that intake and hosting. then other people have bots which download bulk models to collabs everyday. where is all this bandwidth being funded. it's a steep task they're accomplishing. i'm grateful for what it is. but wow is it ever a thirsty crowd.

I have to wonder what the motive for spending so much money is.

vast ingot
#

damn

slender vault
#

Its a shame MJ charges an arm and a leg.

vast ingot
astral goblet
#

i know today in particular something happeend with their image servers. clean up scripts got a little aggressive. they've unplugged a few systems there

vast ingot
slender vault
#

Oh 100% and honestly its great that you do, it helped me get set up allot quicker than it would've on my own

astral goblet
#

MJ is often used as an example of corporate control over AI, but i think we should also recognize that it's a small 6 person team and them keeping their innovative systems proprietary is actually a smart strategy this early in the game

#

i dont always like SaaS, has disadvantages galore. What they've created is impressive though

potent spire
#

MJ is 0 corporation

astral goblet
#

yup

slender vault
#

Like yeah its impressive quality but theres things like Kandinsky coming out thats gonna give MJ a run for their money. SD being open source theres always the chance of someone making things compatible

potent spire
#

If i want corpo which i do, i go to either Open AI more or less or to Adobe

#

My opinion is, for artists Stable Diffusion and Adobe Firefly are the way to go

#

For ready to go, MJ

vast ingot
#

SD being open source makes it the main cement of most new and to come generative picture AI.

#

it's a good thing to build upon, it would be stupid not to use it if you want to go this road

astral goblet
#

sd being open source also allows other research teams to learn from it. kandinsky benefitted a lot from SD existing

potent spire
#

I prefer Firefly especially for the future. But for others Stable Diffusion is the way to go probably

astral goblet
#

corporate and opensource don't always clash either. Meta released that cool new segmenting model yesterday

potent spire
#

With those two you have the most control

#

With MJ you are as of now extremely limited

vast ingot
potent spire
#

So its a ready to go one

astral goblet
#

i'm trying to learn clip and blip right now. interested in training my own image captioning system. those were released by openai and salesforce

potent spire
#

I believe MJ loses long term

#

Even in quality of images

#

Nothing beats the control

astral goblet
#

firefly is a webservice and doesn't run locally if i'm not mistaken. Studio workflows will likely integrate a local install of SD and a plugin instead.

potent spire
#

And skilled artists dont mind repairing images

potent spire
astral goblet
#

Firefly is very much a beta test bed. Whatever the final generative feature set is, it won't have that name

slender vault
#

I cant wait to see if someone is able to make Kandinsky compatible with Auto1111, using just the cpkt isnt enough.

astral goblet
#

Firefly being saas just makes it a non starter for many studios. Keeping everything on internal networks is important because they often don't have the rights to license content to adobe's servers for use in firefly

potent spire
#

When its out of beta i expect the actual explosion of it

astral goblet
#

clients sign over logo masters and branding images, stock photos, etc. you can't use those on a 3rd party service if you don't have the rights to upload that image to them

potent spire
#

For me and many others it will be ideal. Alone because of the native integration with CC apps

astral goblet
#

i don't think adobe's plan is to let it run local

potent spire
#

Local? Nah, they will allow customised model with your own assets

#

But not run locally the whole thing

#

You are bound to their CC likely

astral goblet
#

do you have a source on that? i've never herad them stating they plan to release the model for training

#

customizing a model requires having the model

potent spire
#

Including training on your own assets

astral goblet
#

if they're doing it as saas, that requires hosting all the trained content on their servers

#

if they're doing it as a local install then people will just take the model and implant it into a webui

potent spire
#

Scroll a bit

astral goblet
#

I think the corporate control of ai is going to be a huge effort but it just won't happen

potent spire
#

Personalized results
Generate images based on your own object or style.

Thats what it says

astral goblet
#

all these legal licensing issues are going to stand in their way, and open source is really the only path that's got a future

potent spire
#

They are already very good with a limited dataset...in early beta. They will be huge soon

#

I doubt they will fall here

astral goblet
#

the only way i see it working is if a new copyright act is made to replace the DMCA that creates safe havens for these sort of licensing / hosting content issues

potent spire
#

The amount of ressources they have and the top notch specialised people working for them, puh

#

Thats why i see SD and Firefly as winners in long term

#

And ofc those lawsuits come into the game too, gotta see how this all plays out

astral goblet
#

adobe are leading for a big reason. they have the tech and the results

potent spire
#

And possible regulations

potent spire
#

I dunno how MJ will keep up with those two

#

Although MJ will have an web interface too and will likely apply those controling features too sooner or later

astral goblet
#

it's not so much lawsuits against adobe thats the problem. It's the professional studios, their meat and potatoes, that will have licensing concerns. It's basically industry wide standard practice to get signatures on everything. Hosting content on a 3rd party server gets people in trouble often already, as these very proprietary graphics being licensed out for production reasons shouldn't just be on a random cloud server if the studio doesn't own the rights to pass on to the service provider.

#

saas has a niche market

#

samsung actually had problems with this and chatgpt. and itsa huge problem. clients are going to not want to use the chatgpt api company wide. Samsung employees were pasting very proprietary documents into chatgpt to ask it questions. Samsung wants those documents secret, but their employees were just feeding them to openai servers without even thinking about it

potent spire
#

Lawsuits wont hit Adobe i guess because they have right on Adobe Stock stuff for example

#

They are untouchable in that sense

astral goblet
#

chatgpt is going to hit a hard ceiling because of that. it wont get used as widely as hoped

potent spire
#

not sure about OpenAI and GPT

#

I think they will get away

astral goblet
#

i'm not talking about lawsuits. i'm talking about licensing issues over content between professional studios and their clients. these are the biggest adobe customers

potent spire
#

Aaah okay

#

They are safe as of now

#

And they would be forced anyway to play smarter here and dodge the bullet

astral goblet
#

chatgpt is already a huge success. thats undeniable.

#

companies will be unwilling to integrate it though, since it sends everything back to microsoft servers

#

microsoft sees a lot of success there already with onedrive, but thats changing fast

#

cloud was a fad

void canyon
#

hi i am very happy to join this server.

potent spire
#

@astral goblet Copilot X for example seems promising

#

Its not perfect, but its already used by some companies etc

#

But it wont replace those jobs

#

But ofc lawsuit runs against them too

astral goblet
#

i think copilot will see it's biggest successes in the open source fields

potent spire
#

Id want to give it a try when i get a bit more into coding

#

Maybe i will try it soon actually

#

But at this point i better should monitor my subscription expenses lol

#

Its a madness with me in this case

#

Art alone costs me thousands of € per year

#

Thousands

spark phoenix
#

hi guys, i have a problem when i try to generate an immage with controlnet's tool. It's appears an error message: "RuntimeError: CUDA error: CUBLAS_STATUS_ALLOC_FAILED when calling cublasCreate(handle)".
How can i fix it?

fervent thunder
#

But agan, probably will be good to fix little details or just to type less, but terrible at somewhat complex tasks
(since chatgpt itself is + it will need to be able to read \ remember ALOT more tokens)

gentle prairie
#

so i didn't use stable diffusion (webUI) since last year, what's new and how do i install them? I saw something like T2I or controlnet? Sorry i don't know how any of this works

fervent thunder
gentle prairie
#

huuum

#

i'm doing what it says here :

#

to install from URl

#

but it does nothing

#

when i click install

fervent thunder
#

You can just install it from extension tab

gentle prairie
#

yeah that's what i did, i don't know why it worked after the third try

gentle prairie
#

yeah i don't know what i'm doing wrong with this controlnet stuff

#

when i try to do the same as that, i get just some texture or paint stuff, not even getting a normal image

oak blade
#

Heyy I got a code promo reduction on rundiffusion if anyone need it : misakuara15

gentle prairie
#

also my old styles are broken, i don't know how to load them, it's like it's loading all the styles prompt at the same time.... god why do you have to relearn everything after just a few month, lol

#

what are the best models right now? my last ones are stable diffusion 1.4 and waifu model 1.3

broken smelt
#

was using the standard vae that came with a1111 for the past month 🤦‍♂️
the stabilityai 840k vae model makes some pretty noticable differences

flint raft
#

Can anyone recommend a good model for tie dye? Haven't been able to find one.

narrow quest
#

Can anyone recommend a good solution for referencing model info directly in A1111? Like usage, trigger words, recommended vae, etc.? I'm doing this in Excel rn, and it's painful.

#

I use the civitai.com extension and it rocks, but unfortunately it doesn't pull that data into the .civitai.info file.

digital kraken
#

where do I post images that I've made?

vast ingot
digital kraken
broken smelt
#

idk if its civitai specifically but it looks like the loras that ive downloaded automatically have their keywords on the plugin without me having to type any of them in, pretty neat

#

i read trigger word and i immediately thought of loras so on second thought, mabey it isnt what youre looking for lol

narrow quest
sly eagle
#

Hi!, do you know if it safe to train stable diffusion with a original artstyle with google colab?

manic wraith
#

you mean safe as in if google will steal your data or something?

sly eagle
manic wraith
#

I think google steals all data as a business model, why else would colab be free, but on the upside they steal so much data yours flies under the radar 👍

fervent thunder
#

hello guys

#

stable infusion is completely free for now?

#

how it is financed?

manic wraith
#

it's been free for a long time

#

initially Emad shelled out a metric ton of dollar bills and trained the base model, then other people finetuned it for things like anime, mostly for fun

fervent thunder
#

i have tried it a bit - via webpage - i see there are also forks - forks are using own engine or they still use stable infusion?

manic wraith
#

it's free if you run it on your machine, running on other people's computer usually costs money

fervent thunder
#

sure

sly eagle
#

I just want to make my pics faster 😂

fervent thunder
#

i can share some images I have made however they are nsfw

coral stag
#

Where do we go to diffuse now?
I was only using it during the beta and now all channels are restricted :c

fervent thunder
#

free to use - means some sort of api access?

manic wraith
fervent thunder
#

What is automatic1111?

coral stag
#

How good a PC would you recommend to try and run it yourself haha

fervent thunder
#

I have access to gpu cloud if it helps

manic wraith
#

it has every feature you could think of and then some

coral stag
#

How is it compared to midjourney tho ?

manic wraith
#

harder to use, more control, possibly better results (you can do anything on it), usually worse results if you're not good at using it

#

you need 4GB+ of VRAM

#

ideally more

coral stag
#

Im down for a challange, got 8gb vram
Possible to get a a link ?

manic wraith
#

you can use it for any styles, train your own, etc

#

dunno if it's possible to link it here because the guy had some beef with NovelAI and they banned him from this server (that's all I remember)

#

just search for automatic1111

#

he's on github

coral stag
#

I was never good with github lmao

manic wraith
#

you might be better off just paying for credits on a simplified UI then

coral stag
#

Thanks anyway

fervent thunder
#

is there an ai art server that is platform agnostic? to discuss different platforms and open source tools

#

i am making one myself for fun 🙂

#

just to see how it works

#

Stable Diffusion is a latent diffusion model, a kind of deep generative neural network. Its code and model weights have been released publicly,[4] and it can run on most consumer hardware equipped with a modest GPU with at least 8 GB VRAM.

#

neat

#

@manic wraith are you an artist?

manic wraith
fervent thunder
#

i am single

#

I produce art write books poetry

sly eagle
# manic wraith wife is

Does she hate this? I'm an artist myself but I'm not sure what to feel i just want a pc assistant because I can't draw so fast

manic wraith
fervent thunder
#

I can see that people will pay to listen to a poet live

#

agi in art is fun

sly eagle
manic wraith
#

she's super pro, too, and would wish it wasn't such a drama because then she could offer free BG options or stuff like that

sly eagle
#

I want to train in my own style but not share it publicy

fervent thunder
#

it is better to share it

manic wraith
#

img2img is a cool way to get ideas also when you're stuck

fervent thunder
#

as to model training well I am new to it - I did read on gan big gan

manic wraith
fervent thunder
#

adversarial training and other training approaches

manic wraith
#

there might be a Rentry page explaining the process too

#

anons write those

manic wraith
#

there are other clues from physics to improve the process

fervent thunder
#

Different how?

sly eagle
fervent thunder
#

I think the more AI art is out there the more oil paintings will cost 🙂

#

unless of course they will be painted by ai also

manic wraith
fervent thunder
#

ty

#

The goal is to train a Machine Learning model that learns how to go backwards through time, reversing this corruption process. If we can successfully learn such a mapping, then we have a transformation from a simple distribution (Gaussian, in the case of Diffusion Models) to the data distribution.

#

hmm

#

i am not sure it will work well

#

The laws of physics, therefore, provide an invertible mapping between a simple distribution and the data distribution

#

it says it can avoid mode collapse hmm

molten furnace
#

sits on a stool, cinematic lightning, holding a microphone, looking into camera, photorealistic, detailed face, unreal engine, pixar--q 2

fervent thunder
#

stands on a cliff and looks

#

:)))

hallow plaza
#

Is there a way to blip2 on less vram? Clip inspector is comically bad.

brisk bloom
hallow plaza
brisk bloom
#

I was just linking it as its online so no vram use for your pc

hallow plaza
wanton urchin
burnt lichen
#

Are there any guides out there explaining how to use Stable Diffusion legally for commercial purposes?

wet halo
#

Rule of thumb, is do NOT use key word references to someone else work or source. if you want a picasso, dont say picasso say "infuse geometric shapes into the image" or something like that.

#

The moment you reference someone else's work you hit the legal marker

#

Artists are inspired by other artists.

#

The other thing to note is read the license for the model your using, The general terms allow for commercial use. However some models are specifically trained on an artists source material, if this is the case then it will likely be copyrighted material.

burnt lichen
#

I am literally using a 3D model which I licensed legally, and painted myself in a 3D program, to generate input for img2img. Whenever there's ambiguity, such as whether a streak across the character's neck is a scarf or a choker, I often need to break the tie by editing my prompt to specify what it is. Style is not copyrightable, the pose is mine, the colors and shapes are mine, and the interpretation of what those colors and shapes mean is also my original work. Anything that comes from the training data is a smear of color swatches sampled from, I assume, thousands of images.

#

If I personally did this by hand, using the eyedropper tool in Photoshop to sample pixel colors from other images, I'm pretty sure it would still be transformative to the point of Fair Use.

#

Thoughts?

#

(To clarify, by "input for img2img," I mean the image on the left in Automatic1111. Not input data to train a new model.)

alpine bear
# burnt lichen Thoughts?

take mine with a grain of salt.
The legal argument doesn't matter, especially in your specific case. You're creating a majority of the original material, so no one worth your concern will come after you.

burnt lichen
#

That's what I thought. But I'm listening if anyone wants to play "devil's advocate" and present a counterargument.

vestal dew
#

I agree. The only issue is you cannot copyright AI work, so someone can steal your work.

burnt lichen
#

What about editing in photoshop after generation?

#

For that matter, if I edited the image, it won't even have tags in it anymore. How will anyone be able to prove AI was used to make it?

#

And surely the work my image is used in, such as a game or a movie, is still protected? It's not... an image.

#

At worst, it's a curated anthology. At best, it's a derivative work with its own seperate copyright from the individual output images used to make it.

vestal dew
#

Exactly

#

nobody will know, much less likely care about your work

burnt lichen
#

I just figured the most likely threat would come from a huge corporation suing end-users just for the hell of it.

#

"Time-Warner owns 5% of the images used to train the algorithm you downloaded! Pay up, chump!"

#

You know, business as usual.

#

Then again, I suppose I didn't train the algorithm. I acquired it in good faith.

#

Only a matter of time before Big Media passes new laws to favor themselves as the only ones allowed to use AI , of course. But until then...

tough bough
#

I wonder how effective such laws could even be. The tech is open source and widely used, which means the genie is pretty firmly out of the bottle.

shadow meteor
#

hi

broken smelt
#

i remember reading that it's actually not too pricey (all things considering) to rent the hardware necessary to train a model for a few hours

vast ingot
#

good morning everyone

#

nice interesting questions

#

so to do a new model from scratch, yeah, you'll need lots and lots of data, and computing power. Last person I saw that tried used around 300k photos for the model training
Depending on if you start from scrach or from a base model, more or less training will be required, but you would test this like in any other training : test for bleeding, overtraining and bias

#

yes, 1.5 is the 5th iteration of 1.0

#

Can I create embeddings indefinitely?

#

I don't get that question

#

what do you mean

#

if you are training a new model, you aren't training embeds

#

ho ok

#

concepts

#

not embeds

sly eagle
#

I have a problem that says "AssertionError: Dataset directory doesn't exist" not sure why it happens, I´m using my images folder route

vast ingot
#

on an existing model, you can train as many Text inversion embeds, you can use multiple at once
but only 1 lora embed per prompt (unless composable lora extension)

sly eagle
#

no, i'm running from Colab

vast ingot
#

Yes it does checkpoint. You usually have some params like "do checkpoints every X steps, starting from step Y"
so you can revert to previous states of the training if you overtrain

#

lol

#

that's a lot of steps for 30 pics :p

#

it will be burned x)

#

well depends on the GPU I guess but still

sly eagle
#

no idea what you are talking about, i don't see those options

vast ingot
#

but to respond already

#

overtraining : burned outline, the input pictures in your dataset almost come out of the model

#

bias : some feature come out unprompted, consistently.

#

bleeding : the main concept taught come out on everything

desert umbra
#

Anybody guide me how to create hindu god image in stable diffusion automatic 1111

sly eagle
#

Thank you now i get it sorry to sound so newbie but... i'm new with this waow

vast ingot
#

you test bleeding by running prompts at 0 CFG : if your concept shows, it means it has bled. It needs more regularization data or less training
You test bias by running lots of prompts and seing bias show up. you fix it by hunting in your dataset for examples of this bias in too many pictures, and you remove some
You test overtraining by runing seeds on your main tokens, and see that the outline is burned, or that the face starts to have pixelated artefacts. You fix it by having more diversity in your photos of the concept that overtrained

#

good one already, but if you want to train a full fine tune like this, you are facing multiple days/weeks of continued training from what I got

#

using the right tool for it will be quite essential too, you should try different ones and see what brings you the most img/s trained, on what batch size, ...

#

(nice ELI5 here, well worded)

#

you're a really good explainer

#

how much experience do you have training models for now ?
I would suggest you start with a little smaller project, to expend on your dataset/evaluation capabilities. Like training 10 concepts at once in a model and having all of those respond correctly already may open your eyes to some of the dangers here.
Because you won't be able to test bias easily once you have 1000 concepts for example
Instead, training a smaller model for 20 minutes and learning from it, and then expending to larger models, seems like the way to go

#

yes, that also lets you balance each concept individualy

#

then you can store those 30 pics as an OK dataset

#

and start the next one

#

the goal will be to have lots of OK dataset that train using the same params/total steps

#

so that you can then just put all of those in a single training

#

I mean as a dreambooth, not as a full finetune. Check the guide, I give some tips on the number of pics, but usually, I use around 10 to 15 for a single subject (like a character) and 50 to 100 for a style$

#

yes. tags are very important. The tokens you choose are the ones that get trained. In particular, check what Attention is :

Each step of the training, a batch of pictures is trained, and the weights of the model move a little. Those changes happen slower or faster, depending on the learning rate you use. This "budget" of changes that could happen on a single step is called Attention, and is split amongst the tokens you used in your caption.

Adding more tokens to a caption then has multiple effects :

it slows down the training on each single token. This may require more total steps to produce the same results, or to use more trained tokens at once in your prompt later on.
it looks for the more fitting parts of the picture for that token and associates with it the changes. This means that describing a feature in your caption can prevent that feature from being associated with the other tokens. As an example, if training on a character that has a tie in half the shots, adding the token "tie" would reduce how much of the tie feature is associated with your character.
it spreads the training on more weights of the model, and reduces the need for regularization.

#

I go in great details on the dataset techniques in that guide

#

Token : Common sequences of characters found in text. Usually about 3/4th of a word, they are the parts that constitue your prompt, and the keys to accessing the weights you are training. Choosing a token is choosing a word to train the model on.

#

(I have a full terminology section for all those concepts)

#

bleeding is when your concept starts to be associated with too many tokens

#

it starts to show up whatever the prompt then

#

it's when your dataset isn't precise enough on your concept, or is wrongly tagged, and the training doesn't stick to modifying just the right tokens, it modifies everything

#

well

#

the highest is the best

#

and only depends on your vram

#

highest batch size will do :

  • multiple pictures trained at once, faster
  • better training quality. up to a point, diminishing returns after batch size 6 from my tests
#

yes. testing your tools and taking note of some params are important :

  • the time it takes per image to learn
  • the total time it takes to "cache latents"
#

yeah but on what batch size ?

#

1 ?

#

so 4.3-4.8 img/s

#

this is what you need to keep. it trained around 4.5 img of the dataset per second

#

so if you had 100k picture, each epoch would take 5 days

#

and you usually train for hundreds of epochs

#

(an epoch in this context is going through your whole dataset once)

#

yep

#

for styles, I usually train 100 times per picture, total

#

on 3090TI (what I have), I can train 50 pictures in 10 to 20 minutes, depending on my params

#

you can technically pour as many pics as you want into it :

  • more pics = longer epoch
  • more pics = harder to detect the bad sides of the model (lots more testing)
    but there isn't really a hard limit on number of pics outside your own time
#

with good ventilation you can go for long

#

but for such training... datacenters are better

#

you at least keep your computer to yourself to do other things, and it can even cost less to you in the end

#

going on a 80GBVRAM card can let you go to batch size of 40+ on some tools

#

I upgraded last year, a little after joining the server ^^

#

I almost went for an A100

#

yeah but damn efficient :p

#

I didn't go for it in the end

#

because of the noise

#

and that I didn't want to turn one of my rooms into a refrigerated datacenter

#

3090TI is already a big beast, 24GB

#

let's start a training by the way, I have a dataset ready :p

#

I do models for the server here every week

#

just launched mine

#

here are my models

#

they are trained in a special way I have to say

#

ho

#

that is a little slower imo. I use everydream2trainer

#

"in a special way" explaination

The "classic version" responds to tokens for the 4 elements.

The "user token version" is trained on dedicated tokens per user, and doesn't work like you may be used to.

I try to make a model out of all the submissions, for people to continue enjoy the theme after the event, and see a little of their designs in other people's creations. The token stays "SDArt" and I balance the learning on the low side, so that it doesn't just replicate creations.

The pictures were tagged using the token "SDArt", and an arbitrary token given to the user that submitted it.

#

all the Automatic other dependencies overhead takes lots of vram and doesn't let you go as high in batch size, so it's slower

#

this is deterministic. most tools do have a "seed" for training too

#

This is what one of my training params file look like

  "amp": true,
  "batch_size": 4,
  "ckpt_every_n_minutes": null,
  "clip_grad_norm": null,
  "clip_skip": 0,
  "cond_dropout": 0.04,
  "data_root": "F:\\AI\\Data\\Datasets\\Preparation\\CoW\\prepared\\futuristic clothes",
  "disable_textenc_training": false,
  "disable_xformers": false,
  "flip_p": 0.0,
  "gpuid": 0,
  "gradient_checkpointing": true,
  "grad_accum": 1,
  "logdir": "logs",
  "log_step": 25,
  "lowvram": false,
  "lr": 1.5e-06,
  "lr_decay_steps": 0,
  "lr_scheduler": "polynomial",
  "lr_warmup_steps": null,
  "max_epochs": 100,
  "notebook": false,
  "project_name": "SDArt_FuturisticFashion",
  "resolution": 512,
  "resume_ckpt": "F:\\AI\\Data\\Diffusers\\stable-diffusion-v1-5",
  "sample_steps": 100000000,
  "save_ckpt_dir": null,
  "save_ckpts_from_n_epochs": 100,
  "save_every_n_epochs": 20,
  "save_optimizer": false,
  "scale_lr": false,
  "seed": 555,
  "shuffle_tags": false,
  "useadam8bit": true,
  "validation_config": null,
  "wandb": false,
  "write_schedule": false,
  "rated_dataset": false,
  "rated_dataset_target_dropout_percent": 50,
  "zero_frequency_noise_ratio": 0.02
}```
burnt lichen
#

Can you clarify what it would be an infringement of? style is not protected under U.S. law. It's not protect-able. Any artist is allowed to ape any other artist's style. Just not their intellectual property or their expressions.

vast ingot
#

automatic uses one of the base dreambooth implementations, adds LoRA into it too if you want, and works on ckpt mostly
The tool I use doesn't work exactly the same, there isn't "reg data" at all, it's a big soup of all my dataset using the file names as prompt for the corresponding picture, and works on diffusers directly

white herald
#

hi

vast ingot
#

.....

vast ingot
# white herald hi

hi and welcome but you'll need to change username to stay around here mate 😉

#

I changed it for you for now, sorry

#

ok training has ended

#

let's test it :p

#

using automatic, if you interrupt, it should make a save point at the step you are at

#

it will resume, but it may not be 100% the same as if you had just ran it continuously

#

in particular, very small changes making it not deterministic to reproduce, in the order the images are trained on

#

it seems really long already, you have a good card, doesn't it has created some checkpoints already ?

#

I would test what it gives right now already yeah

#

training for a full day is no way to learn

#

I'm just testing the 32 tokens I trained

kind kraken
#

what is the difference between Inpaint area and Mask mode ? I sort of understand what Mask mode means, basically inpaint whatever is masked.
But what does Inpaint area means?
It has like two options
Whole Pictture and Only Masked

vast ingot
#

this option is basically how much context you give to the AI too

#

nice !

#

it/s will be slower, but each iteration is of the size of the batch

next haven
#

Went in for the first time. There's so much here. Where can you ask questions? 🙂

vast ingot
#

tagging purple eyes isn't always the best imo. I tend to opt to tag less and not more

vast ingot
#

and we have lots of dedicated chanels for specific problematics

#

but the general chat is cool too

#

we can't post pictures in this current channel though

next haven
#

Prompt tab Image browser. Is it possible to increase the number of simultaneously viewed pictures from 36 to 50, for example. And how photos get into the Favorites tab.

vast ingot
#

It was slowing down too much my automatic this extension, I removed it. But it should be possible I imagine.
Such settings go into the file "ui-config.json" usually, if the extension creates such option

#

you can modify all sliders of the UI in this file

#

(it requires to restart automatic for the changes to happen)

#

I'm not 100% sure it's available for that extension though

next haven
#

Thank you.

vast ingot
#

I don't see it in there :/ (i installed the browser to test)

#

nice, did you vary on the background too ?

#

nice

#

should work. I wouldn't have removed the glass though but yeah

#

not sure what you mean there. Don't use transparency though, it's rarely giving any good results, mostly bad

next haven
#

Tell me how to use Artist to study in Stable Diffusion. I can not understand. It seems to be absolutely simple. But where is the button "apply" style, I can not understand.

vast ingot
#

it's not through studying artists though

#

it's just making prompts for you

next haven
#

Thank you.

vast ingot
#

seems on the high side of things. Unless you are a transformer, some of those may be too many.
Having lots of pictures, when the face stays the same in it, may bring some difficulties :

  • you will need more total steps for the training to understand all pictures, so it may start to hammer the rest of the model too much : regularization may become more needed
  • some details may repeat more than others and overtrain faster. Imagine if your nose overtrains but not your eyes, because removing the glass in half the pics made the eyes train slower for example ? This will happen a lot less on a lower step count, so less pictures can be better
#

as for the caption, both schools work, but I prefer the second one, it has less chances to fail
1/ first school is using a blip caption, and replace the "a man" by your choosen token, and fix some captions that bugged.
2/ second school is to tag all your pics the same : with only your token

#

having the same tag on all pictures is quite strong. as long as the dataset is diversified on what it presents, it will only learn what is static : you. not the color of your shirt or the background

#

initialization text is what to put inside your token at the start of the training. You can put a prompt in there, it will use the weights to initialize your embed

#

it brings it faster to the good result to initialize with a good description of yourself, your ethnicity, ...

#

put either "a man"/"a woman" or a longer description of you in there

#

and for the vectors per token, this represents 2 things :

  • how much your embed can "store" data. the more you put the more details it can have, but also require more training
  • how many tokens in your prompt will this embed cost to use
#

usually for a single character/person, we use 5 to 8 tokens

#

do your hair change in the pictures ? ^^

#

do your glass change in the pictures ?

#

because adding those tokens to your prompt will make it so :

  • that the tokens "short hair" and "wearing glass" will more resemble YOU
  • that your token on its own, without "short hair", may make other hairs
#

so I wouldn't specify the hairs at all in the prompt

#

those are just "you"

#

same for the glass

#

you can specify the shirt and the background for example

#

but don't put any of those 3. There is not really a use for those. And I'm not sure about the convention itself

next haven
#

@vast ingotAre you the one who developed the automatic 1111? Please accept my respect!

vast ingot
#

lol nope

#

it's AUTOMATIC#1462 that did develop this

#

I'm just a moderator and community member here

next haven
#

@vast ingotThank you all. It really changed my life for the better 🙂

vast ingot
next haven
#

I wonder how the program will change in 8 months? Will I understand everything 🙂

vast ingot
#

none at all

#

no hairs, no ears, no nose, nothing

#

only the things you don't want learned : shirt, background