Automated layout generation from sites to flats using GAN and transfer learning

Abstract

Generating architectural layouts from sites to flats, encompassing site layouts (SLs), building layouts (BLs), and flat layouts (FLs), presents a complex process. Notably, the BL generation is challenging due to the small scale of data, making it difficult to train effective neural networks. This paper introduces an approach for generating layouts throughout the complete process. Initially, it proposes an enhanced generative adversarial network (GAN) combined with the transformer for Stable Diffusion (TranSD-GAN), considering design boundaries and requirements. Subsequently, for generating BLs with small-scale datasets, the paper proposes a stacking transfer learning method. Following this, image operations are conducted to support the flow of building information. Ultimately, BIM models are created at each stage. Through comparative experiments involving neural networks and generation cases, it is demonstrated that the proposed method significantly improves the generative capabilities of small-scale datasets and effectively aids designers throughout the layout design from sites to flats.