If I understand correctly, they use both in parallel: low resolution, animated voxels for the light map, and high resolution, detailed rasterized polygons for t

If I understand correctly, they use both in parallel: low resolution, animated voxels for the light map, and high resolution, detailed rasterized polygons for the final rendering pass.