Given a corruption process whose likelihood can be evaluated and differentiated, the authors model the prior and posterior distributions by Normalizing Flows and jointly train them by optimizing the ELBO.
In one of their experiment, they pre-train the prior distribution on source data and the posterior distribution on pairs of sources and corrupted data. Then, they solve the optimization problem. The training objective improves but the prior distribution get worse because the variational bound is not tight.