One thing that this work requires is some from of symbolic differentiation. Basically, given some expression e, you have to compute, via a symbolic analysis on e's structure, a derivative for e, e'. e' can then be used to compute normals used for shading, replacing the need for a hand crafted normal map (you can also approximate normals via sampling, but if you have a functional definition of a surface, why not just use derivatives!).

This definitely requires some compiler support, but the purity of Elm should be helpful in that regard.