Democratizing the power of urban space visualization through inclusive AI tools.
Creating visual representations of urban public spaces requires specific skills, which can influence the development of city environments. Our objective is to make these visualization tools accessible to all community members, especially those who are often underrepresented. By providing these tools, we aim to allow more people to participate in decisions about the spaces they use.
To support this, we developed the Mid-space dataset. This dataset includes 3,350 prompts, 16,694 images, and over 42,000 annotations. The annotations are based on six criteria: accessibility, safety, diversity, inclusivity, invitingness, and comfort. These criteria help in evaluating different aspects of public spaces.
In this page, we describe how the Mid-space dataset was created, examine the annotations collected, and briefly show how the visualization tools can be adjusted to fit community needs. Mid-space combines artificial intelligence with urban design to incorporate input from community members into the development of public spaces. This project provides urban designers with data that reflects various community values, supporting the creation of public spaces that consider diverse perspectives.
The Mid-space dataset offers a wide range of public space ideas, represented through numerous Stable Diffusion XL generated images and detailed by extensive community annotations. These annotations cover varied community preferences, crucial for studying accessible and equitable urban spaces.
Mid-space was developed in collaboration with Montreal’s community organizations, including the Congolese Community Center of Montreal, Altergo, La Maisonnée, Cummings Centre, Projet Changement- Community Center for Seniors, Women’s Center of Plateau Mont-Royal, LGBTQ+ Community Center of Montreal, Marguerite-Bourgeoys Hub, RÉZO, Afrique au Féminin, L’Agence On est là!, and Montreal Women’s Groups Table. This partnership employs a hands-on approach by involving local residents in every stage of the process, from defining evaluation criteria to reviewing images. This ensures that the dataset authentically reflects the diverse urban needs of the community.
The development of Mid-space followed a structured process to accurately reflect community needs. Starting with initial workshops to identify what makes spaces inclusive, the project advanced through prompt creation, image generation, and detailed annotation, illustrating a method for developing responsible AI and responsive public spaces.
The above figure shows how fine-tuning the Stable Diffusion XL using Direct Preference Optimization affects image outputs. Comparing images of a meditation garden—original, inclusivity-optimized, and fully criteria-optimized—highlights how the model understood specific community preferences, leading to more relevant and functional urban spaces.
For more information, contact: rashidmushkani@gmail.com | www.rsdmu.com