How do they guarantee that the learnt marginal vector field will transform the distribution such that when sampled later it creates sensible data samples?