[moultano]

That's not really true. The residuals of a good lossy compression algorithm ought to look like uncorrelated noise with a smaller dynamic range than the original image, and the second fact (smaller dynamic range) makes them compressible.

[DarkShikari]

Dynamic range is not really relevant to compression. Entropy is much more meaningful, and error images typically have a huge amount of entropy.

[moultano]

That doesn't make any sense. You've taken an image with arbitrary bytes, and turned it into one where the bytes are tightly centered around a small range of values.That's perfect for Golomb coding (for instance.)

[jerf]

No free lunches. For a lossless compressor, (Information theory bits of polygons) + (Information theory bits of remaining error) >= (Information theory bits of original image), where the > represents the possibility that the first two elements aren't perfectly separated by your process or have inescapable overlap.

I specify "information theory bits" because they aren't really what you see in the computer's RAM; they're closer to "post-compression bits". But regardless, no matter how you move the encoding around there is no escaping information theory.

[moultano]

Obviously. That's the definition of compression. If the image is well modeled by overlapping polygons + small residuals, the encoding will better approach the uncomputable ideal, and thus, compress.