generative building facades realizes architecture renewal in hong kong

Generative Building Facades by Jianing Luo and Boyuan Yu

Hong Kong’s rapid urbanization has resulted in a complex mixture of historic and modern architectural styles, often causing visual identity discord across city blocks. Jianing Luo and Boyuan Yu’s project, Generative Building Facades with Pix2Pix GAN in Hong Kong, responds to the pressing challenge of harmonizing new development and renovation efforts with existing neighborhood character. They propose a design methodology using Pix2Pix GAN — a deep generative model — to automatically generate building facades that are contextually aligned with Hong Kong’s urban environment.

the workflow from the input to the generation of output

The research behind Jianing Luo and Boyuan Yu‘s Generative Building Facades with Pix2Pix GAN in Hong Kong, begins by constructing a custom dataset of 160 facade images from old neighborhoods in the city. These images, collected via Google Maps and archival photography, were manually processed into bitmap label maps that define key architectural elements — such as windows, balconies, signs, and walls — using a simplified color-coded system. Each image was aligned, normalized, and resized to 256×256 pixels for optimal training conditions.

the color categories in the building label image

Pix2Pix Generative Adversarial Networks

The core of the design process utilizes conditional Generative Adversarial Networks (cGANs), specifically the Pix2Pix architecture, which includes a U-Net-based generation tested on 32 pairs, using data augmentation techniques to enhance robustness. Testing revealed that the AI-generated facades demonstrate high fidelity to Hong Kong’s vernacular forms. In particular, the system successfully reconstructed elements like varied window arrangements, floor divisions, and street-level shopfronts.

data augmentation

While reflections and weathering in real facades posed challenges, the generated outputs preserved core compositional logic and provided cleaner, stylized alternatives, potentially useful for early-stage design explorations and facade design in dense urban contexts. It not only aids in preserving the architectural heritage of historic districts but also reduces the subjectivity and labor intensity of traditional facade analysis. The trained model can serve as a generative design aid for architects and planners, offering a fast, scalable, and data-informed means of producing facade proposals for urban renewal and infill development. Future work will expand the dataset, refine labelling methods, and explore integration with 3D scanning technologies to enhance precision and application.

generator construction

testing results

training process of generator and discriminator

the generator and discriminator loss values

existing buildings and automatically generated buildings output

project info:

name: Generative Building Facades with Pix2Pix GAN in Hong Kong
designers: Jianing Luo and Boyuan Yu

designboom has received this project from our DIY submissions feature, where we welcome our readers to submit their own work for publication. see more project submissions from our readers here.