In this work, we propose a vision-based and Finite Element Method (FEM) based controller to automate the 3D shaping of soft objects with dual-arm robots. Our controller relies on a data-based approach to learn how the robot's actions result in object deformations, while also running FEM-based simulations to infer the shape of the whole body. These model-based simulations are used to generate initial shape data, allowing to extract visual features through a Principal Component Analysis and thus estimate the interaction matrix of the object-robot system. In contrast with most existing shape servoing controllers, our new model-based approach continuously predicts the object deformations produced by the robot, which are then compared to the visually observed deformation feedback. This iterative process enables to correct the deformed mesh model before updating the interaction matrix. To validate this new control methodology, we present a detailed experimental study with a dual-arm robot and different soft objects, which showcases the performance of our automatic shaping framework.