Because the rapid development of World Wide Web technology and the advances of computer hardware and software, a better visualization and presentation of three-dimensional map has become possible. The large amount of data volume of three-dimensional map, however, will lower the performance of map display. For example, visualizing three-dimensional building model will normally encounter the conflicting problem on data quantity and display performance. Data generalization not only simplifies the detail of the features presented on the maps, but also provides the superior visualization effect in different map scales. Many studies have used the concept of LOD (Level of Detail) to simplify raster data, which can reduce and remove the redundancy of the display data effectively, and achieve the goal of the high performance and quality. However, very few studies have focused on the generalization of three-dimensional data in vector form. In this study, we have applied the concept of LOD and have designed a procedure to process the multi-scale generalization for three-dimensional building vector data in urban area. The method deals with the problem in three particular steps: (1) to clean and build road network into block polygon; (2) to cluster, aggregate and simplify the building polygons in each block polygon; (3) to determine the multi-scale LOD thresholds for the generalization. As a result, we have developed five thresholds to convert three-dimensional building polygons into five hierarchical layers. Accordingly, we have designed and developed an internet service for the LOD vector data provider and used the Google Earth for the client platform to display the three-dimensional building model in urban area. The successful operation of multi-scale generalization of three-dimensional vector data shows that major application of the proposed method is in the field of the real-time internet transmission and visualization.