The advances in the automatic acquisition of 3D geospatial data has lead to large-scale urban landscape models that can easily consist nowadays of thousands of building objects. For a photo-realistic visualisation, each building model has a number of highly detailed façade textures associated with. Because of the vast amount of objects and façade images, a brute-force rendering approach is not feasible for such scenes even on high-performance visualisation systems. In this paper, two methods are presented to improve the rendering performance of 3D city models. The first approach is a generalisation algorithm that automatically generates simplified versions of 3D building models that can then be used in level of detail structures. Compared to already known surface simplification algorithms from the field of computer graphics, the presented solution yields more sophisticated building models by combining least squares adjustment theory with an elaborate set of surface classification und simplification operations. The concept allows the integration of surface regularities into the building models which are important for the visual impression. These regularities are stringently preserved over the course of the generalisation process. The second approach is an image caching technique that accelerates the rendering process by reusing parts of previous rendered images in the current frame. Using this so called impostors technique, real-time rendering of large-scale urban landscape models is possible nowadays even on low-cost PC hardware equipped with commodity 3D graphics accelerators.