1: \begin{abstract}

2: Barycenters are  principled summaries of populations of measures. To estimate barycenters, we typically parametrize them as weighted sums of Diracs and optimize their weights and/or locations. This approach, however, does not scale to high dimensions due to the curse of dimensionality. In this paper, we propose a technique for facilitating difficult barycenter estimation problems through a different parametrization of the barycenter by means of a generative model. This turns the barycenter estimation into an optimization problem over model parameters, which sidesteps the curse of dimensionality and allows for incorporating inductive biases directly into the model. We prove local convergence under mild assumptions on the discrepancy, thereby showing that the approach is well-posed. We demonstrate that our method achieves good performance on low-dimensional problems and provide unprecedented results by scaling barycenter estimation effectively to high-dimensional image data. 

3: \end{abstract}

4: