So I've been playing with Stable Diffussion again. in particular the Automatic 1111 program on my desktop which is not at all suited to my almost 2 year old laptop so is very slow and is glitchy as all get out but it's been about 6 months since I last played around with this and I have to say the improvement in what it can do since then is truly astounding. It feel to me a little like the start of digital photography, early days it was kinda useless and then all of a sudden it was amazing.
So I present here three photos that I quite like out of the machine. I've found a 'model' I like which is Hyper Photo realistic, which can produce a pretty real looking person, but then I found what I enjoyed more was using prompts which give me something a little different
This prompt is simple, 'A couple in a Dsytopian street, one is dressed as a rabbit the other a clown." I enjoy that it kind of gets it but also kind of doesn't, but it works.
This is 'A creature who is half woman, half tiger.
And 'A rabbit'
One of the things I find really interesting is about the bias in the learning of AI,
In the first shot all I said was couple, and I got a male, female mix. I never specify age but they are young, same with race, but I got white. Super interesting, none of the prompts suggested any of this should be sexual in any way but I have a male fully dressed and a female in a skimpy outfit, a naked Tiger lady (I guess Tigers don't wear clothes so maybe that makes sense) but why is the a 'rabbit' not a rabbit but a female robot rabbit, or maybe it's a female rabbit lady in armour. Maybe it's the people playing with this maybe it's the effect of the images it's learning with
And this is the concern with AI, much like the algorithms in social media, will it magnify the prejudices that are already in society, or media and will we need to actively push against them