Nice. Not dissimilar from what jpeg2000 does (from memory - correct me if I’m wrong!). They use the previous pixel values in order to “guess” the next value and then store the delta, which tends to be a smaller number on average, giving better compression.
proximity in a 2D image has neighbors in north-south-east-west (NSEW) on a 2D grid, not just pixel neighbors per-scanline (row of bytes); proximity in a sound file has harmonics and transitions of some plural set of signal, but the sound file neighbors are roughly at a point in time. corrections welcome
True. The fact of locality in real-world 2D images means you can predict a pixel's value by picking any neighboring pixel, or indeed any nearby pixel, with worsening performance with L2 distance. There's no reason I can think of to prefer a direction, except for convenience in the data structure.
