Evolution of implicit inference posteriors according to the number of simulations used to train the NF. The posteriors (blue contours) are approximated using the NLE method with simulations only. We train 5 NFs with the same architecture where only the initialization of the weights of the NF changes. Each blue contours correspond to the approximated posterior of one of these NFs. The ground truth (black contours) corresponds to the explicit inference posterior of 160 000 samples and the black marker corresponds to its mean. The yellow marker corresponds to the mean of the approximated posterior. With few simulations (e.g. 100 simulations) every NF predicts a different posterior and each prediction is overconfident. With a bit more simulations (e.g. 1 000 simulations), the posteriors are consistent and similar to the ground truth.