Nowadays, the extraction of buildings from aerial imagery is mainly done through deep convolutional neural networks (DCNNs). Buildings are predicted as binary pixel masks and then regularized to polygons. Restricted by nearby occlusions (such as trees), building eaves, and sometimes imperfect imagery data, these results can hardly be used to generate detailed building footprints comparable to authoritative data. Therefore, most products can only be used for mapping at smaller map scale. The level of detail that should be retained is normally determined by the scale parameter in the regularization algorithm. However, this scale information has been already defined in cartography. From existing maps of different scales, neural network can be used to learn such scale information implicitly. The network can perform generalization directly on the mask output and generate multi-scale building maps at once. In this work, a pipeline method is proposed, which can generate multi-scale building maps from aerial imagery directly. We used a land cover classification model to provide the building blobs. With the models pre-trained for cartographic building generalization, blobs were generalized to three target map scales, 1:10,000, 1:15,000, and 1:25,000. After post-processing with vectorization and regularization, multi-scale building maps were generated and then compared with existing authoritative building data qualitatively and quantitatively. In addition, change detection was performed and suggestions for unmapped buildings could be provided at a desired map scale. . © 2020 International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences - ISPRS Archives.