What: local AI image generation.
Why would AudioSex care: album covers, images for wavetable synthesizers.
How: demons in your GPU, probably.
What you need: Windows/Linux, 16GB of RAM, an NVIDIA GPU with (ideally) 8+GB of VRAM, around 20GB of storage minimum

NO, YOU CAN'T GET GOOD HANDS OR TEXT OUT OF THE BOX (YET)

Windows:​

1. Install Git for Windows and make sure to check the box for the Windows Explorer integration.
2. Right click anywhere, select 'Git Bash here' and execute
   Code:

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

3. Download a model - this is what the AI will use to generate images (these weigh up to 8GB each)
   Stable Diffusion 1.5 (all-purpose but outdated), Deliberate (all-purpose, realistic), Dreamshaper (oil paintings, fantasy art, scenery), more can be found here (though you should use this filter)
4. Place the model(s) you downloaded in the '/models/Stable-diffusion' folder
5. Install Python 3.10 and choose 'add to PATH' - the Windows 7 installer can be found here
   IT HAS TO BE THIS SPECIFIC VERSION because python sucks ass
6. (Optional): for a sizable speedup and VRAM usage reduction on Pascal (GTX 10xx) cards and newer, open up 'webui-user.bat' and add '--xformers' after 'COMMANDLINE_ARGS='
   FOR 4-6GB OF VRAM, ADD '--medvram' and expect it to take a while
   FOR <4GB OF VRAM, ADD '--lowvram' and go make some coffee
   FOR GTX 16XX CARDS, ADD '--precision full --no-half' in addition to '--medvram''
7. Launch 'webui-user.bat' (NOT as an administrator), the first launch might take more than 15 minutes
8. Visit 127.0.0.1:7860 on your browser of choice

Linux:​

1. Install python3 through whatever package manager you use ('wget git python3' for Red Hat and Arch-based, 'wget git python3 python3-venv' for Debian-based)
2. Run
   Code:

   bash <(wget -qO- https://raw.githubusercontent.com/AUTOMATIC1111/stable-diffusion-webui/master/webui.sh)

   in your folder of choice

• Run 'webui.sh'

Recommended settings
Start the WebUI, go to the "Settings" tab and do the following:​

• 'Stable Diffusion' tab: enable 'Make K-diffusion samplers produce same images in a batch as when making a single image'
• 'User interfaceì tab: paste 'sd_model_checkpoint, sd_vae, CLIP_stop_at_last_layers' in the quicksettings list box
• 'Live previews' tab: DISABLE 'SHOW PROGRESSBAR' AND 'SHOW LIVE PREVIEWS OF THE CREATED IMAGE'
• 'Sampler parameters' tab: set ETA noise seed delta to 31337
• 'Actionsì tab: request browser notifications (useful)
• CLICK THE BIG ORANGE APPLY SETTINGS BUTTON, CLOSE THE CMD WINDOWS AND RESTART THE WEBUI

How to:


How to get bigger images on SD1.5-based models:
WIP, will write later​

 

Last edited: May 2, 2023

Did not try this yet, but thank you in advance for your time/ work.

Two things come to mind - can Python 3.10 be installed on W10? Can the 8GB models reside on i.e. 2nd SSD?

Thanx

 

Yeah, the first link is for Windows 8 through 11, the second one is for Windows 7 only. And yes, you can store them on a second drive. There's a way to change where it looks for models but I don't remember where that setting is stored - you can move the entire "models" folder somewhere else and symlink it back to where it was.

 

That's a lot of hoops to jump through... but thanks!

 

Nah not really, most of your time will he spent waiting for it to launch for the first time

 

I am going to try this today.
The way I interpreted all the AI web tools fine print is they also own anything you upload or make with them.
I have a gtx1070 and a rtx3090ti -- will be interesting to see how they perform.
Thanks for taking the time for all this information.

 

Most online services "own" what you make, yes. Not sure if this applies to the paid ones too like NovelAI and Dreamstudio.
In some ambiguous SD benchmark I found the 3090/Ti (the difference between them in SD is even less than the 10% difference in gaming) is like 9 times faster than the 1070 Ti (which is a little over half as fast as the 1080 Ti)

 

Thanks Xeraser
Great instructions - working on windows 10
Did not take long at all.
I think it is time to start learning python because I have no idea what just happened lol.

 

Think it just took forever on mine due to my shitty 11 years old i5, good to know it's not supposed to take that long.
You don't need to learn python for this, there are dozens of extensions available for this specific UI and even idiot-proof training scripts in case you want to add new data (images or even new concepts) to the models (though you might need all 24GB of VRAM depending on how you want to do it)

 

says xformers unavailable
when making 6 images it takes about 15 seconds with CPU (Ryzen 9 5800) and Video not going above 10% usage
also with it open python is using 5gb of ram if anyone was curious of resources used
any recommendations for websites to learn more?
thanks again

 

...or you can take a look at this.
https://aescripts.com/diffusae/
If you are using After Effects, this is really an nice and easy approach.

 

xformers not found on a 10xx series card or newer? That's odd, it should automatically download a prebuilt xformer library when using a compatible GPU
Why are you using your CPU though? It shouldn't do that by default, check your GPU usage while it's generating.
You can find some artists to prompt ("by" followed by the artist name) here Comparison of all Artists in Stable Diffusion (urania.ai)

It's a paid plugin (both prices are ridiculous) that's EXTREMELY limited, there isn't a single good reason to consider it.

 

Last edited: May 2, 2023

You can also give https://nightcafe.studio/ a try if setting this up looks too daunting. It has stable diffusion and much more! People are making some super detailed and amazing creations. It helps to use starting images as well to get more in line with what you're trying to create.

 

Services like that work on a credit basis, sooner or later you'll have to pay up to keep using it and you're limited to what they offer in terms of features and models. If you have the hardware required to run this smoothly I wholly recommend it over using one of those services. Plus these services usually "claim" whatever you generate. Not an issue if you're using the images as inputs for wavetable synths but if you're trying to sell an album...

 

it hovered around 10% use at idle it stays at 2%
so it did use some

 

If you have an iPhone 11+ or M1 iPad/Mac you can try Draw Things: AI-assisted Image Generation
Heats up iPhones quite a bit but iPads cope well with it.

 

Last edited: May 9, 2023

Ok Audiosexualism, I am here, I will try to do this. Thanks to OP and all for sharing.