3D city models play an important role in multiple applications, but creating them still requires effort using various possible techniques. This paper proposes a new machine-learning-based framework for generating 3D city models. With the help of conditional Generative Adversarial Networks and single orthographic images, segmentation and height estimations of buildings are achieved. The height information per pixel and the building coordinates were generalized using a histogram for heights and the Douglas-Peucker algorithm. The framework was evaluated by using variations of the same dataset (for the city of Berlin) to show possible differences due to changes in the image size and representation of the heights. The evaluation reveals that it is possible to generate block models with a mean absolute height error of 5.53m per building, a mean absolute height error for the whole raster of 1.32m, and a Jaccard Index of 0.55 for the segmentation. While the proposed framework for generating LoD1 city models does not attain the accuracy of previous techniques, our work represents a step towards successfully using machine learning for the automatic generation of city models and building segmentation.