This is a simple script to find a personality value (PID) that most closely matches a target image when applied as a pattern to the Pokémon Spinda!
You should go see the original, this is just a tribute. I was inspired by adef's vid, of course, and the spirit of procrastination. All credit for the real work goes to them, I'm just playing with the result.
This version has been optimized for speed a bit, most prominently by replacing the cool genetic algorithm with a much less cool but ~60x faster call out to scipy's bounded minimize (L-BFGS-B I guess?). I had to use a custom Jacobian since the loss function between original and "Spindafied" image is discrete (can't have a spot between pixels!). I've learned that apparently optimizing a discrete function, even an approximant of a nice smooth one, sucks a lot! So the solver struggles: while the results are faster, they lack the same visual fideility ("Spidelity?") as the original. Doesn't stop me from doing crimes:
I've also introduced some incompatibilities, or I'd just make this a PR (json -> numpy output of PIDs, removed some features, and I wreaked some true havoc on large_spinda.py). I'm happy to work to upstream this stuff if there's suddenly a community of like-minded Spinda-pattern sickos that would want an alternative "Spindafiability solver."
However, after actually learning how Spinda patterns work, I'm sure there are much more dramatic optimizations than this: don't bother solving at all, just find an approximate inverse function! You could find particularly dark/light locations (maybe with edge detection?) and then translate their centers or borders into suitable spot locations, with the additional hard-coded "white" and "black" available for removing spots that aren't needed. While technically the spots can overlap for some PIDs, they don't really need to in order to replicate (almost) any pixel configuration, so there should be a unique satisfactory configuration available just from arranging the spots to follow the clearest light/dark edges.
There's also a lot of room to optimize images for Spindafication beforehand: thresholding and compressing them more intelligently to 1-bit BW might help both the program speed and the visual readability of the final Spindafied version (not least by reducing solver failures, hopefully).
File list:
spindafy.pydefines theSpindaConfigclass, representing a Spinda point configuration. It also contains a method to render the pattern to a PIL Image, and to compare it to a target image.large_spinda.pyimplements the main mosaic generator, splitting an image into zones and giving each a Spindafied replacement.large_spinda_anim.pymakes an animation oflarge_spinda.pyresults.spinda_optimizer.pyis the untouched original genetic algorithm version, for comparison.