Got talking with my tiki advisor and newly excited for that novel. So many basic questions to solve yet. Take world-building. What kind of world am I after? What kind of world-building am I using (that is, how much of the Gernsbakian explain-the-tech stuff is appropriate for the style I'm after)? How much world-building at all?
I mean, I don't even know if there are aliens.
Minor work on my personal tiki shrine -- aka, things I can have around me while I write. Found a nifty (plastic, but it looks okay) mp3 player shaped like an old transistor radio. Have tiki mugs. Thinking rattan mats...maybe even a palm.
***
Also retro is the new LORA I tried out in my SD install. Broke Automatic111 again (which is good) but finding it easier and easier to do a clean re-install so it is working again (which is bad). I'm only doing stuff for my own amusement and all I need in my life is another time sink.
But anyhow. I'd been bookmarking some "look like an illustration" LORA but not using them because I didn't quite see the point. Now I do. There's an unreality about AI creations which is rather off-putting. Well, turns out that making it look like an illustration and not a photograph eases past the worst of the Uncanny Valley.
That, and this library is just really deep, and seems very good at interpreting. That could be the new model; tried out SDXL, now using a fork off that. It is much better at parsing prompts, though things like color words do have a tendency to go wild and spam over everything (the "man on a beach with a white shirt" problem).
I was having very little luck txt2img with the new LORA and was about to give up on it. See, most illustrations have story in them -- even if the "story" is just "proud engineer points at the new car." And I am all about story. Well, AI can only associate pixel patterns. That's one of the basic problems. There's no underlying logic outside of "this pixel pattern is often found near this pixel pattern" (which is why a wire will turn into a panel line half-way across a control panel, and vice-versa).
And there's no understanding of function, which is not just "the wire needs to be connected at both ends" (and not turn into a painted border in the middle), but the underlying logic to, say, a pose.
But it can inherit that given a proper reference. Turns out between the model and the LORA my local SD install has become really, really good at "reading" a source image and reconstructing what is actually going on. Well, things like humans holding swords or which part of a torch is actually on fire, it still has trouble with. But enough of the original logic of the source image is getting preserved that there is some decent mimicry of, well, story.
***
And speaking of time-sinks. Still feeling so lousy that Satisfactory is all I feel up for some days. My last video just reached 1K views and a bunch of likes, which is a huge surprise. I guess people like my aesthetic.
But I am contemplating another build, this time heavily modded and with most of the alternate play options switched on, like free materials and no spiders.
One of the mods allows you to replace your trains with more historical-looking ones. With that and a texture mod, could maybe perhaps do something that isn't modern brutalism-plus-neon (which is practically the default look for most of the extreme builders. Called by one of my video commenters "reminds me of Portal."
No comments:
Post a Comment