Given the fact that this data chunk uses Predictor function 12, we can see on pg52 of pdfReference 1.6 that this corresponds to PNG prediction (on encoding, PNG Up on all rows)
In the middle of page 51 of the reference doc, we can see that the PNG predictor functions are defined within RFC 2083.
Here's the relevant section: RFC 2083, PNG - Section 6 Filter(ed: predictor) Functions
It may be worth a look inside the code of pngLib or some other source-code distribution of a PNG writing/reading library, just to see how the algorithm has been successfully implemented.
I'd certainly be most interested in any/all progress. I'm Implementing a pdf creation class in c++, and have just abandoned lzwDecode owing to the lack of documentation on how to go about creating a PDF-legal lzw stream. Have found a couple of micro zlib implementations - typically only implementing the deflate method and any pre-requisites.
So, this is a little further up my alley than I am currently, though do hope to squish every last bit out of any encoded streams - it seems I'll need to use PNG predictors too at some point.