Ai, just wanted to say I admire your AA research!
I only hope the results won't be so good to tempt lazy pixellers 
(just kidding, of course)
Did you just apply it on your avatar, or was it already like that ? heh.
Thanks
Here's an update.. I've changed my method again and it's much simpler (and WAY more effective
)
* Use GMIC to apply a Difference of Gaussians on a copy of the source image,
to detect areas of high frequency detail like eyes (parameters: 0.05, 0.10, threshold 20, enable 'negative'.
* Apply a curve with a point at 93,33 and 192,192 to reduce the amount of smoothing high-detail pixels will get, even more.
* Edit->Copy
* Enter QMask mode (using keyboard shortcut or the button at bottom-left of your window) and Edit->Paste what you just copied.
* Anchor the pasted layer and exit QMask mode. Now, the low-detail pixels are most selected, and high-detail are least selected.
* Apply GMIC Anisotropic smoothing (amplitude 8, sharpness .4, gradsmooth 0, tensor smooth 0.50, iterations 2)
* Done
Here's the previous attempt, for comparison:
And here is one I tried today:
(done with slightly different smoothing parameters -- sharp 0.30, iterations 1, to get stronger lines.)
Some things to note:
* outer lines are preserved, inner lines are AAed. So we don't sel out.
* Face details still end up smoothed somewhat more than I intended, but
I view this as an acceptable tradeoff -- a few touchups of detail vs manually AAing *everything*. complicated shapes such as hands seem to be preserved *quite* well
* you might want some more AA on long lines.. I've concluded the most useful
approach is either to AA them manually, or select and smooth them separately to the main iteration, using GMIC Anisotropic Smoothing In that case you mainly care about sharpness (which controls amount of smoothing relative to amount of color difference), amplitude (which can be quite high for much more smoothing if you set sharpness well.. if not, soggy result), and possibly secondarily gradient smoothness, which effects how smooth borders are interpreted as being, and tensor smoothness, which effects the smoothing of the angles in shapes
* extremely high contrast areas are left alone (which is generally good, unless you actually DIDN'T want your high contrast to stay high)
* Depending on the exact style of the pixel art, sharpness will need adjusting.
I'm quite happy, the major remaining challenge is to auto recolor back to plausible colors found in the original graphic.
(other things: try the process on other graphical styles, eg. SNK)
The results are good enough (with high amplitude value) to help as a general guide of how you could AA a shape well. Hopefully it would be used only as a 'finishing' filter
in all other cases.
