cade.site ⇒ Guide to Stable Diffusion

About | Blog | Links | Art | Timeline | Archive

Guide to Stable Diffusion

image thumbnail

Stable Diffusion, an open-source AI text-image model, was recently released. It’s notable for being the first “big-boy” (i.e. high fidelity, state-of-the-art quality) model that is completely freely available.

If you haven’t already, check out my post about DALL-E 2 for some background on what the heck this AI art stuff even is. If you’re lookin

Setup

Free Demo

There’s a free demo on huggingface, which is a great way to test it out. Although it can be ran from anywhere (even mobile), it’s slow since it’s running on a free tier, open queue.

Good for just expirementing but not for serious work.

Google Colab Notebook (web client)

If you don’t have or want to use a GPU or hardware to run the model yourself (or you’re not a nerd who can follow programming instructions), you can use the Google Colab notebook. This is a web-based client that lets you run the model in the cloud. It’s free, but you’ll need a Google account to use it.

There’s a few different notebooks out there, and will probably be more and more made. You should google “Stable Diffusion colab notebook” if these are out of date.

Local: UNIX setup for lstein/stable-diffusion

If you use an Apple Silicon Mac, you can use lstein’s fork of Stable Diffusion, which adds support for M1/M2 macs, as well as a very cool web interface that makes prototyping easy.

Just follow the setup instructions; they work as is. However, some of the scripts had problems running on my M1 macbook, so I just used the dream web interface instead:

# run the web interface
$ python3 scripts/dream.py --web

Usage

In this section I’ll share some tips and tricks for generating.

Prompts

Cyberpunk ethereal alien planet with rainbow mist auras, dense mushroom forests, canyons, and river deltas. 8k high quality digital art landscape

Ancient caverns of existence with a 7 dimensional god meditating with a rainbow aura. 4k digital art and oil painting

Results

Check out all my stable diffusion results here. Here’s a sample:




theme: