A FRAMEWORK FOR FINE-TUNING URBAN GANS USING DESIGN DECISION DATA GENERATED BY ARCHITECTS THROUGH GANS APPLICATIONS

Recent studies have utilized Generative Adversarial Networks (GANs) to learn from existing urban layouts for urban design tasks. We define these GANs as Urban-GAN. However, urban layouts generated by Urban-GAN lack specificity and often require multiple modifications by architects to meet specific design requirements, making the process inefficient and non-customizable. Inspired by the concept of fine-tuning language models, we propose a stacked GAN model framework that fine-tunes Urban-GAN using data generated by architects in solving specific design tasks, forming AD-Urban-GAN. Our results indicate that layouts produced by AD-Urban-GAN more effectively emulate architects' design morphology decisions, enhancing Urban-GAN’s adaptability and efficiency in handling design tasks. Furthermore, AD-Urban-GAN enhances the customizability of Urban-GAN models for specific urban design tasks, generating layouts that accurately understand and meet the requirements of specific tasks. AD Urban-GAN significantly streamlines the process of generating design prototypes for specific task types, enabling precise quantitative control over urban layout results. This workflow establishes a data acquisition and training loop that strengthens the customizability of existing GANs. The design decision data generated by architects can improve the adaptability and customization of GANs models, facilitating efficient collaborative work between architects and artificial intelligence.