[partyvan]

hoo whee. ok so:

• using edge reflect could segfault
• euclidean was meaningless because hits were already getting clamped to a circular radius (this managed to elude me for months :DD)
• aa was basically doing the opposite of what it was supposed to so i Crafted some Science: img below is the difference between a full-res precise distance function for a couple of shapes, and a --distance-reduce 4 --distance-spread 2 of a half-res antialiased shape, @ 8x bilinear:
  • 
  • aa actually helps a bunch now so at least for the past 5 months it was the thought that counted :')
• dither did ~nothing outside of a very particular set of images; practically everything about it was wrong and the few textures in my game looked better with it on basically by complete accident because the quadtree was basically one big FIGUREITOUT leaf so the indices weren't completely wrong (but even then it was diffusing error in the wrong direction oopsie)
  • it's roughly on par with a naive floyd-steinberg so hard to say where the value is besides maybe being faster cause it skips a bunch of pixels but hey it's got a unique character to it
• speaking of faster it goes way faster now thanks to this paper: https://cs.brown.edu/people/pfelzens/papers/dt-final.pdf
  • basically turns the distance transform separable, which is cool except it also discards a bunch of hits, so adapting dither/AA, now that they're Good Actually and not Hot Garbage, was a bit of a pain
    • super extreme reductions with AA are kind of unavoidably slow cause it does still bail out to a manual scan when it didn't get anything close enough from the transform; tying it more tightly/""""elegantly"""" to the transform would basically i think amount to running the transform over a supersampled coverage estimation? cant think of a way youd make that fast on cpu but theres a fun google summer of code project for someone
    • practical settings like --distance-reduce 16 --distance-spread 4 still go basically as fast as plain when you turn on aa, and are noticeably higher quality than default

[craftablescience]

Holy shit, awesome, thanks!

[partyvan]

i leave the particulars of updating language binding docs about the two flags to you, as you've done before, but i've put in a maretf pr that i think says everything it needs to, when maretf gets this sourcepp revision