How to Make AI Generative Design for Building Façade Renovation More Controllable: A Case Study on a Historic District in Zhejiang, China
This study proposes a dual-control generative design framework integrating Stable Diffusion, ControlNet, MLSD, and LoRA to balance historic preservation with commercial revitalization in Zhejiang's historic districts, achieving high structural fidelity and functional responsiveness for façade renovation across diverse business formats.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the bustling streets of China's historic commercial districts, a quiet tension plays out between the past and the present. These neighborhoods, often built along canals with white walls and black-tiled roofs, are not just museums; they are living spaces where shops, cafes, and homes must adapt to modern life. The challenge for city planners and designers is to update these buildings so they can host new businesses without destroying the historic character that makes them special. If every shop owner renovates their storefront independently, the street can quickly become a chaotic patchwork of clashing styles, losing the visual harmony that defines the area. To solve this, researchers are turning to a new kind of digital tool: generative artificial intelligence. Unlike older computer programs that simply follow rigid rules, this technology can create new images from scratch based on descriptions. However, standard versions of this technology often struggle with architecture, producing buildings that look beautiful but have impossible structures or misplaced doors. The goal is to find a way to guide this creative power so it respects the bones of an old building while fitting the specific needs of a new business.
A team of researchers from Hangzhou City University has developed a method to make this process much more reliable, specifically for the historic districts of northern Zhejiang province. They created a system that acts like a double-check for the computer, ensuring that any new design it proposes keeps the original building's shape intact while also placing functional elements like entrances and windows exactly where they are needed. The researchers tested their approach on four common types of businesses found in these areas: restaurants, retail shops, coffee houses, and guesthouses. By combining a technique that locks in the building's structural lines with another that maps out where business features should go, they were able to generate renovation plans that look realistic and respect the local architectural style.
The process begins with a photograph of an existing building. The computer first analyzes the image to identify the permanent parts of the structure, such as the roofline, the main columns, and the overall proportions of the wall. It strips away temporary details like specific doors or windows to create a clean skeleton of the building. This skeleton is crucial because it prevents the computer from inventing new roof shapes or changing the height of the floors, which would be impossible to build in a real historic district. Once this structural frame is locked in, the researchers overlay a simple color-coded map that represents the needs of a specific business. For a restaurant, the map might highlight a large area for a display window and a clear path for an entrance. For a guesthouse, it might mark out a balcony and keep the walls more solid for privacy. These two layers—the unchangeable structure and the flexible business needs—are fed into the artificial intelligence model together.
To ensure the generated images look like they belong in northern Zhejiang, the researchers also trained the system on a specific set of local architectural photos. This step teaches the computer the subtle details of the region, such as the texture of weathered white walls, the pattern of black tiles, and the way light falls on timber frames. When the system generates a new design, it uses the structural skeleton to keep the building standing true to its original form, the color map to place the business features correctly, and the local training to apply the right materials and atmosphere. The result is a series of renovation proposals that look like they could actually exist on the street, rather than just being artistic fantasies.
The researchers tested their method by generating designs for the four business types and comparing them against what a standard computer program would produce without these specific controls. They found that without the structural guide, the computer tended to change the roof shapes and column arrangements, creating designs that would require tearing down the original building. Without the business map, the computer would often place entrances in the wrong spots or forget to include necessary features like balconies. When both guides were used together, the designs stayed true to the original building's shape while successfully adapting to the new function.
To verify how well this worked, the team asked a group of architects and design experts to review the images. The experts judged whether the generated designs kept the original building's structure and whether they met the functional requirements of the business. The results showed that the dual-control method achieved a structural accuracy of 77.6 percent and a functional success rate of 75.6 percent. This means that in roughly three out of four cases, the computer produced a design that looked structurally sound and functionally appropriate. The team also measured the visual quality of the images using standard computer vision metrics, finding that the generated pictures were visually plausible and consistent with real-world examples of these historic districts.
The study highlights that this tool is best used as a way to visualize ideas early in the planning process, helping authorities and business owners agree on a direction before any physical work begins. It is not yet a tool for creating the final construction blueprints, as the images are two-dimensional and lack the precise depth information needed for building. However, it offers a powerful way to balance the preservation of history with the demands of modern commerce. By giving designers a way to test many different business scenarios on the same historic building without losing the street's character, the method provides a practical path forward for revitalizing these cultural treasures. The research suggests that with the right controls, artificial intelligence can become a partner in preserving the past, ensuring that historic streets remain vibrant places for both residents and visitors.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.