~insom — now


   2023-01-21
   ----------

   Hi. No one needs to read this but I want to get the thoughts
   out of my head on AI generated art. Feel free to skip.

   I'm specifically talking about Stable Diffusion, which is
   the only thing I've really used, and that I have a local
   copy of, so I'm not limited by cloud computing credits or
   however people are limited on the public / proprietary
   models.

   I'm not like, a _booster_ of this tech -- it's going to take
   another already underpaid field and make it worse, and in
   many cases used a bunch of artists' work uncompensated to
   help to kneecap their field. Pretty dark.

   That said, the technology itself is amazing that it works at
   all. And it's probably not going away now. I feel like we
   can wait out blockchain until it dies off, but as AI art
   generation can be used to make labour cheaper for capital
   it's almost guaranteed to stick around.

   I get that it's fun on social media to joke about how it
   can't count fingers, but the real answer to that is just ...
   scale. I have an RTX3060 video card and I can generate ~200
   images on a text prompt in a couple of hours, and I can
   train a model on around a 1000 images to a reasonable degree
   in about 12. Running things locally changes the dynamics
   entirely:

   Yes, many images are unusably bad, ranging from "a little
   off" to "many mouthed eldritch horror" but when all you've
   done is sink a dollar's worth of electricity into something,
   only one of those 200 images needs to be usable. It turns
   the process into curating and selecting from a large body of
   possible images. Also: retouching and cropping still exist,
   so even if you end up with one of those 6+ fingered hands,
   it doesn't render the whole thing unusable.

   IMO, when the marginal cost of creating things through
   automation drops so close to zero, humans end up getting
   fucked. Self checkouts are bad, but they let one person
   supervise 6 lanes, so companies put up with their badness.
   Google Translate is highly flawed but it's free for most
   uses and real translators are (justifiably) expensive. AI
   images look weird but even cheap illustration is orders of
   magnitude more expensive. People will lower their standards
   -- very far -- for something that's free or cheap. I'm
   guilty of this too -- I've literally ordered two or three of
   the same part from AliExpress in the hope that at least one
   of them will actually be real (and not a fake or a clone) --
   it's still often cheaper than buying one of that same part
   from Digikey.

   On a personal note: as someone who has tried on and off to
   get better at drawing, knowing that this technology exists
   is extremely demotivating. Even if I put in regular practice
   for the rest of my life, it's very likely that someone with
   a dollar worth of GPU time will always be able to create
   something which _looks_ better than I can produce. I used to
   just be comparing myself to artists (which is, itself, a
   negative habit) -- now I'd be comparing myself to _anyone
   with a tool_.

Tilde Town