In case the artist doesn't draw the X,Y, and Z surface normal components directly but instead chooses some other set of lighting profiles, you could use photometric stereo to recover the surface normals. (If this is the approach used, then applying such a technique to specially-crafted pixel art is indeed novel).
Here's a factorization technique for photometric stereo that could be applied to the artist inputs:
I do a lot of research work with stereo, so thanks for the link! Have to give that a go sometime :)