Mixture Manifold Networks: A Computationally Efficient Baseline for Inverse Modeling