I recently started playing with Stable Diffusion, just to see what all of the fuss was about. I mean, I KNOW, what the fuss is about, but I wanted to see how easy it was to create something with it and see how viable it was for any sort of mock-ups.
Here’s a first attempt with a Maciste-related prompt.
Here are some choice results from a raygun adventure-related prompt.
As you could see, those results were terrible. So, I wanted to see if, at bare minimum, I could get it to generate a touhou.
One thing I realized was that I had to set inference steps waaaay above the default. I got substantially better results at 70-90 range + adding in an artist to “train” it.
Vermeer with 70 steps. Still pretty oof.
Renoir with 70 steps.
Renoir with 90 steps.
So, here’s my thoughts on Stable Diffusion. It reminds me a lot of the junior officer in Critical Point: she’s the only one on the station who can get the food printers to put out more than just unappealing space mush because she knows all of the various input configurations, levels and settings [ironically, she’s actually a terrible cook]. It may not be “art” but there’s certainly an artistry to it.
I understand why many artists find the rise in AI art worrisome, and the latest scandal involving DeviantArt default opting in its users’ works to train AI raise some serious red flags in ethics and handling of this new technology.
However, I’m seeing firsthand the weaknesses of the tool for commercial purposes. Yes, it’s proven its worth as a cute girl generator [despite my own failures to output anything particularly worthwhile], but its results are unpredictable and can’t deliver consistency necessary for the kind of work we need. It was a fun experiment, and it’s a neat toy. But we never had any intention of using any art but that created by professional artists.
If you ARE serious about trying out Stable Diffusion to create artwork for concept pieces, Rawle Nyanzi has put together some pretty good tutorials and shared his experience with it.