I've been reading a bit on probabilistic programming, and one of the main claims is that it is more expressive than graphical models. As the representational capacity of PPLs is anything that can be written in the form of compositions of (possibly stochastic) functions, this certainly makes sense. For instance, one can formulate discriminative models and not only generative ones in PPLs.
But what about advantages in representing generative procedures? That is, what's an example where a generative procedure cannot be described by a graphical model?