Why Stable Diffusion Might Be The Future Of Blog Illustration

2022-08-29 18:41:24 PDT (last update 2022-09-06 21:56:47 PDT)

I wrote a couple of days ago about my use of DALL-E for blog illustrations. My conclusions weren't so positive.

Today I discovered Stable Diffusion, an open(ish) source project in the style of DALL-E that you can run on your home computer.

Of course I did. The thing has only been available for a few days, and already there are reasonably easy ways to use it. My NVIDIA 2080 is barely powerful enough to run Stable Diffusion's generator: but it works.

Here's my reaction, in parallel to what I wrote previously.

• On Business Models and Creativity… Being able to run SD at home means I just don't have to think about whether I want to generate an image. It's about 20 seconds per generated image on my box, which is fine.

• On The Limitations Of 2022 AI… SD is quite comparable to DALL-E. So far it seems to do better in some ways and worse in others.

  The thing everybody's excited about with SD is the unclamping of limits on what can be generated. I guess that matters: I can now ask for a picture of whatever celebrity and get them. Meh.

  One thing that is cool about the interface to SD that I'm using is that it allows "image to image" transformations with text guidance. So starting with my FOB logo image

  

  I asked for it as a Mondrian

  

  and as an Edward Hopper

  

  These were reasonably successful. Other experiments went pretty badly.

• On GitAtom's Clunkiness And "Ease-Of-Use"… Because of the way the SD interface I'm using locally generates images, I can probably include them here a little faster and easier. But that is offset by the hassle and time of starting up SD in the first place.

• On The Competition… No change. It's still just an alternative.

So… yeah. Progress marches on too fast to count.

Dinnertime. Peace, y'all.