normalizing flows are generative models that provide tractable density
estimation via an invertible transformation from a simple base distribution to
a complex target distribution. However, this technique cannot directly model
data supported on an unknown low-dimensional manifold, a co