I think the “smallest” is the interesting part of this exercise. I think it needs to aim for lossless, with a sliding scale of quality that loses detail to the point where a human would notice.
Consider an SVG from [Primitive](https://github.com/fogleman/primitive) and a mostly-transparent overlay containing the diff against the original image. Tuning Primitive is not entirely deterministic (like your algorithm) and I think a human interface for the front-end would allow a few knobs to erase parts of the diff in order to shrink the detail imperceptively. Seems like a good opportunity for a contest :)
2 comments
[ 292 ms ] story [ 438 ms ] threadLossless seems hard, but since the concept of 'seed function' is open for everything. It could perhaps use tricks used in other lossless formats.
> Seems like a good opportunity for a contest :)
Primitive seems to be winning for now :D as far as quality of the images :D.
It has some tricks to decide the size, color, shape of the 'seed' it places. At the moment, my algorithm does not really put any effort into that.