I have invented a bold and revolutionary method in order to compress any picture to only a few bytes (typically more or less one hundred).

Here is how it works in a few words :
It separates the three layers of a picture (RGB, but only stores one specific bit if they are similar).
For each layer, it defines the average brightness.
It divides a picture into a grid, where each cell is either "more light" or "more dark" than the parent average brightness (each is coded in one bit).

Then it makes two average pictures out of the "light" and "dark" areas, and repeat the operation on them.

It stops doing this when the picture is smaller than one pixel or the color is solid.

The source code for this is there

This is of course a lossy method, here are a few examples of original and compressed pictures (typically to about 128 bytes in my own .bpc format, converted back to .png here) :
A picture of Abraham Lincoln :

A baloon :

A camera :

A wooden house :

A flower :

Some album cover :

A typical landscape :


And here is an example of .bpc file. It could probably fit in one tweet.


I didn't know about it back then but it might have something to do with the concept of self-affinity.


Here is an ulterior somehow similar test using voronoi shapes and quadtree based compression.