While definitely cool, I wonder how effective this is. Code/random binary data should have different characteristics than image data.
Each of the possible filters is very simple, but the difference between a great PNG exporter, a mediocre one, and a trash fire is the use of a heuristic to decide which of the pre-processing filters should be applied to each line of output. libpng, the free implementation, includes a heuristic that does a fairly OK job, if you just don't implement a heuristic and use no filtering at all, the results are enormous PNG files as seen in turn of the century Adobe Photoshop. So, no, a general purpose compression algorithm isn't very good for image data on its own.
(Choose for yourself whether you think Adobe wanted to discourage use of a popular free image format in favour of licensed formats for which it held relevant IP, or their development team are just incompetent morons, or both).
Yes, as a means of making the data fed into the compressor repetitive enough to benefit from the compression algorithm.
> So, no, a general purpose compression algorithm isn't very good for image data on its own.
That's not what I said.
You definitely won't save enough bytes with better input heuristics than the amount of code needed to decompress them if you're only working with 13Kb of data.