Yeah, my technique was equivalent to filter mode 0 (the only mode), filter type 1 (sub pixel to the left). literally array.ravel().diff(prepend=0) in numpy.

I think part of the reason png is so slow is that it empirically tests to find the best one.

> Compression is further improved by choosing filter types adaptively on a line-by-line basis.

I'd be curious to unpack a png and see the statistics of how much each kind is used.