Physicists routinely need probabilistic models for a number of tasks such as
parameter inference or the generation of new realizations of a field.
Establishing such models for highly non-Gaussian fields is a challenge,
especially when the number of samples is limited. In this paper, we