The actual representation is using voxels to index a spherical harmonic coding of scene detail (I think), and that includes transparency so the novel view synthesis accretes light along the ray length. This lets it represent transparency and other lighting effects that aren't possible with simple photogrammetry.