MID-Space: Aligning Diverse Communities' Needs to Inclusive Public Spaces

Shravan Nayak, Rashid Mushkani, Hugo Berard, Allison Cohen, Shin Koseki, Hadrien Bertrand

Keywords: Preference alignment, Text-to-Image, AI in urban planning, Community-driven AI

TL;DR: We introduce the MID-Space dataset to align AI-generated visualizations of public spaces with pluralistic human needs and preferences.

Abstract: The ability to create visualizations of urban public spaces is a unique skillset that confers disproportionate power and influence over the city's architectural outcomes. Our goal is to democratize that power; putting easy-to-use visualization tools in the hands of marginalized community members so that they can expand their influence over the spaces they occupy. Furthermore, we aim to finetune these visualization tools using images that align with localized notions of equitable, diverse and inclusive public space. To achieve this, we built the MID-Space dataset. It contains preferences for urban public spaces based on criteria such as inclusivity, diversity and comfort. In this paper, we discuss our dataset development process, analyze the annotations obtained and demonstrate the potential for aligned models.

Submission Number: 50