A two-pass paletted pixel-scaling algorithm that uses weighting counting of adjacent colors and a fitness function (for tie-breaking) to create a 2x scale image.
This is not an efficient implementation, just a quick-and-dirty proof of concept. So it is mainly useful for offline rendering right now, but a few optimizations to create less temporary memory and it could be made pretty quick. In particular, the best_sample function will create a dictionary every call, resulting in a lot of garbage. This algorithm could directly work on an indexed image instead and then the weight array be a fixed-length array that is the size of the image color palette (possibly 16 or 256-color or whatever) that's shared between calls and just cleared before use, and then this should result in way fewer allocations. Also somebody could write it in a systems language like C++ or Rust instead of Python -- which would also help a lot, and hopefully wouldn't be too bad to port.
Tu