Abstract: This paper proposes a new efficient modeling and real-time rendering approach for large-scale urban scenes. Our approach hierarchically recovers the geometry and texture of building units from their aerial and ground-level images, to construct an urban architectural model library. Besides, it sets up a library of layout templates to support the stylized design of each district in the urban scenes efficiently. By incorporating model simplification and paging scene culling techniques into the rendering pipeline, we successfully achieve the real-time rendering of large-scale complex urban scene under rain, snow and other weather conditions. The average rendering rate of the whole scene reaches up to 20 fps.