Interesting questions — in terms of size I would say that the bigger the size, the smoother the final result, but that’s not necessarily the interesting part of the technique.

It should work with portraits too. It all depends on how clear the facial landmarks are so that the extraction algorithm can easily and accurately pinpoint them. So I suspect it will work better with hyperrealistic portraits than with Giotto or Picasso. Colour/monochrome should not affect it, while good/bad lighting (or low/high contrast) will affect it more.

The key thing to understand is that accuracy for these techniques is not a binary property, but a linear scale. Also, that similar datasets will yield better result than dataset with high variance of lighting or technique, which is why the MP photo dataset is a really good one: same photographer, same setting, same lighting.