This thesis explores uncertainty quantification in simulation-based inference (SBI) using deep generative models. By focusing on probabilistic frameworks, the study addresses the inverse problem inherent in SBI, where traditional likelihood functions are often intractable. Four generative models are examined: two Conditional Variational Autoencoders (cVAEs), one Masked Autoregressive Flow (MAF), and one Neural Spline Flow (NSF). These models are evaluated for their effectiveness in learning benchmark posterior distributions. While individual assessments are conducted, all models are compared using the Maximum Mean Discrepancy (MMD) metric, with the MAF model demonstrating the best overall performance. The work concludes by suggesting future research directions involving alternative generative models and evaluation metrics.