Just take the small paragraph about that upsampling. The idea is relative straight forward. When you have generated content that is based on coarse material that was previously hires and therefore downscaled, you are looking for an inverse projection, that transforms the coarse data into hires data. Bilinear upscaling is OK, but it doesn't take edges and surface normals into account. So you compare hires and coarse surface normals (and depths) per pixel and calculate sampling weights for a bilateral upscaling.