Variational inference is a great approach for doing really complex, often intractable Bayesian inference in approximate form. Common methods (e.g. ADVI) lack from complexity so that approximate posterior does not reveal the true nature of underlying problem. In some applications it can yield unreliable decisions.
Recently on NIPS 2017 OPVI framework was presented. It generalizes variational inverence so that the problem is build with blocks. The first and essential block is Model itself. Second is Approximation, in some cases $log Q(D)$ is not really needed. Necessity depends on the third and forth part of that black box, Operator and Test Function respectively.
Not only ADVI and Langevin Stein Operator VI are applicable with OPVI framework. Normalizing, Householder Flows fit well for it.