If the compression can take into account the sub-manifold of potential outputs that people would actually be interested in watching a movie about it can achieve enormously higher compression than if it doesn't know about this.