In recent years, there has been a strong interaction between deep learning theory and partial differential equation theory. In particular, in the field of generative modeling, deep learning models using continuous equations, called Flow Matching (FM), have attracted much attention. In this talk, we will present an attempt to apply FM to conditional probability distributions. Experimentally, it will be shown that the approximate solution of a Dirichlet problem that takes values in Wasserstein space enables robust generation with respect to the condition.