This chapter presents two contributions to the SBI methodology applied to hierarchical models, which have few ”global/population” parameters and numerous ”object-specific” parameters. The first concerns the strategy and particularities of efficient inference of all unknown parameters using a single neural network. The second targets cases in which the number of observed objects (and hence the number of their object-specific parameters) is a-priori unknown: so-called catalogue-based inference, strongly advocating for the use of a set-based neural network with flexible-cardinality input.