Generative Image Models for Evaluating Architectural Design

The illustration shows two silhouettes of cities. In the middle, there is an arrow pointing from the left silhouette to the right silhouette. — © MPI for Human Development

© MPI for Human Development

We spend most of our lives in and around buildings—in our homes, workplaces, and public spaces. These environments are more than just functional; they can make us feel safe, relaxed, or at home. But what exactly makes a space feel safe, inviting, and homey? And how do different architectural features influence our perception and emotional response?

We explore these questions by combining insights from environmental psychology, architecture, and computer science. Our goal is to better understand how people evaluate the design of buildings and urban spaces and to uncover which visual features—such as color or layout—play a key role in shaping these impressions.

To do this, we use generative image models, particularly Generative Adversarial Networks (GANs), which have become increasingly powerful in recent years. These models are trained on large datasets of architectural images and can generate entirely new, realistic visuals. The advantage of using artificial images is that they allow us to control and manipulate specific features systematically, something that's difficult with real-world photographs.

In a series of studies, participants are shown both real and AI-generated images of houses and asked to rate them based on qualities like hominess, invitingness, relaxation, or safety. By analyzing how different visual changes affect people's evaluations, we can begin to map out the design elements that matter most to human perception.

Ultimately, this research aims to yield a deeper understanding of how design elements affect our experience of spaces. These insights could support architects, urban planners, and designers in creating environments that better match our needs and well-being.

Literature

Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., ... & Bengio, Y. (2014). Generative adversarial nets. Advances in neural information processing systems, 27.

Pohlmann, K., Tawil, N., Brick, T. R., & Kühn, S. (2024). When houses wear faces: Reverse correlation applied to architectural design. Journal of Environmental Psychology, 98, 102401.