The representation of uncertainties in numerical weather prediction (NWP) relies on an ensemble approach, where multiple trajectories of the atmospheric system are calculated in parallel using physical models of the atmosphere. This ultimately allows estimating the distribution of possible future outcomes. However, the cost of the temporal integration of NWP physical models limits the size of operational ensembles, thereby restricting the exploration of trajectory distributions. In this context, generative AI models offer a promising avenue to enhance ensembles at a lower cost by quickly generating new members. This is partly the focus of the ANR POESY project (2021-2024, Principal Investigator: Laure Raynaud, CNRM). Our objective is to use generative deep learning methods to produce meteorological fields similar to those derived from the AROME physical model, which is operationally used at Météo-France, and thus enrich the forecast ensemble. In this talk, a brief overview of the research activities conducted during this last two years in the context of the ANR POESY project is given. We will show that generative models can be skilled at generating physically coherent multivariate fields. These generation capabilities can be conditioned on the days’ forecast, effectively generating new members and allowing to better characterize the distribution of possible futures outcomes. The quality of the enriched ensembles is assessed using traditional scoring rules such as the Continuous Ranked Probability Score (CRPS) and Brier score, along with other metrics such as rank histograms and reliability diagrams.