This paper analyses the performance of two of the most well known deep learning-based point cloud coding solutions, considering the training conditions. Several works have recently been published on point cloud machine learning-based coding, following the recent tendency on image coding. These codecs are typically seen as a set of predefined trained machines. However, the performance of such models is usually very dependent of their training, and little work has been considered on the stability of the codecs’ performance, as well as the possible influence of the loss function parameters, and the increasing number of training epochs. The evaluation experiments are supported in a generic test set with point clouds representing objects and also more complex scenes, using the point to point metric (PSNR D1), as several studies revealed the good quality representation of this geometry-only point cloud metric.