Mason's asking about what you changed, but honestly the real win is that "melted toy" stage. That's the model actually learning something weird from your data, not just failing. I got decent results once I stopped trying to fix every bad output and instead saved a bunch of those messed up images. Looking at what the model was getting wrong told me more about my dataset problems than the good outputs ever did.

Man, that hits home. Used to be all about saving the shop time on jobs like that. Then I saw a repair I did on a Camaro's rocker panel come back two years later. The filler had cracked right along the weld seam and it looked awful. Ever since, if it's on a hard line or a high stress area, the whole panel goes in. The extra time sucks, but the comeback is way worse.

Had a similar thing with an old laptop screen. The exact panel was discontinued, so I used one from a newer version that had the same connector. Had to trim a tiny bit of plastic from the bezel, but it's been my daily driver for two years now with zero issues. Sometimes you just gotta make it work.

Ugh, that's sunk cost fallacy everywhere.

The US patent office opened in 1790. They gave patent number one for a potash process. So yeah, by 1843 it was a normal system. Checkers isn't cheating, it's just old. The real surprise is how fast they set up that legal stuff after the revolution.