Schematic of the training of the normalizing flow. The encoder is the pre-trained network illustrated in Figure B.1 and the normalizing flow consists of 6 bijective transformations f_φi where the i^th transformation is parameterized by φ_i, which represents the weights of the MAF layer containing (32, 32) units with tanh activations followed by a permutation layer. Additionally, a batchnorm layer is added as the first layer within f_φ1.