First, One image One time inserted for training. Second, One image One Hundred times inserted for training. It didn’t take long to break things apart.
From there I made some progress considering how the system might interpret two images which are similar but different. The predominant shape being this window (it is seen from the inside on one piece and outside in the other.)
What interested me about the Duck Hunt image and it’s reinterpretation/degradation is how the system latched onto the key details. Pattern recognition. So I wanted to push that with this training. I inserted 50 copies of each of the images above with a visual/formal description of each. Working intuitively from here I had fun generating with random prompts that would try and insert themselves into the predominant shape and/or the predominant details: the siding, the rooftop texture, the trees, etc
It also frequently outputted decorative frames unprompted. Which is interesting how the system might be associating this shape with the broader data set. Really fascinating. But as always, this is leading to the next experiment I’m working on now.
To recap, I think it’s clear that the Inside/Outside training worked because it included formal elements which had detail. This is really apparent in any images generated using that system. Which I think is also why the silhouetted images did not work. There was simply nothing for the system to grab onto. This is all speculation, of course.
Much of my work is about failure, so I kind of don’t mind leaving things off in that state. So much of ‘fine art’ is about mastery, but ultimately 99% of a practice is the failures to get to the 1%. (Or at least my practice is...).
Inside/Outside was a really delightful thing to stumble upon though. I think what the system latched onto is what I ultimately love about this duo of photographs - just how symmetrical they are. Conceptually they touch on a lot of excitement and anxieties I have about technology which allows us to be so hyper-connected. Formally they (in my opinion) are just so nice to look at. So I leave you with some more images I’ve generated using those two images as the training.