AI's learning curve from text prompts, at least as far as images-making is concerned, proceeds rather like an Internet shopping algorithm: "Other people with similar tastes to yours purchased tan pants so you will want to buy tan pants also." In other words, AI interprets what it thinks you want based on what other people have wanted in the past, which probably isn't exactly what you want. It adds it's own AI gloss to your most specific prompting. For example, the image below has a certain Steampunkish tone, which wasn't what I intended at all.
What was I going for? Well, I suppose an Edward Burra or Rene Magritte type figure, in Boston, in the twilight of his career salvaging legacy media from the scrapheap of history. Instead, there's an ambivalent avian Private Investigator from some sort of pre-war science-fiction comic strip, rummaging through microfilm of top secret prototype bird-feeders stolen from the vaults of the Customs House, in a race against time before he must fly south for the winter.
No doubt AI will keep learning, just as the Amazon algorithm will one day work out I prefer grey to tan pants, despite the fact there's a guy in Wisconsin with a standing order for tan pants who also bought the same Last Year At Marienbad DVD that I did. The funny thing is, none of the characters in Last Year At Marienbad would be caught dead in tan pants. But it's going to take years before AI can assimilate such fine shade of niche detail.
Comments