A promising direction to go would be to generate gradients using the polygons instead of colors directly and use Poisson blending [1,2] with gradient mixing [e.

A promising direction to go would be to generate gradients using the polygons instead of colors directly and use Poisson blending [1,2] with gradient mixing [e.g. 3] to generate the final image from that. It would lead to many fewer noticeable artifacts at a similar compression ratio. But I should point out that, while a fun hack, this is not really a viable compression method in practice. [1] Original paper: http://scholar.google.com/scholar?cluster=7038222778464379047&hl=en&as_sdt=1,48 http://scholar.google.com/scholar?cluster=703822277846437904... [2] Some easier-to-understand slides about it: http://www.cs.unc.edu/~lazebnik/research/fall08/jia_pan.pdf http://www.cs.unc.edu/~lazebnik/research/fall08/jia_pan.pdf [3] Someone's project implementing it: http://www.cs.brown.edu/courses/csci1950-g/results/proj2/edwallac/ http://www.cs.brown.edu/courses/csci1950-g/results/proj2/edw...