Co-Creating the City We See: Aligning AI to Reflect Diverse, Intersectional, and Place-Specific Visions of Inclusion

Pluralistic alignment for cities: six criteria, many voices, no one-size-fits-all.

Read on arXiv · OpenReview (ICML 2025) · Project & Dataset

The Big Idea

LIVS (Local Intersectional Visual Spaces) is a people-in-the-loop benchmark for multi-criteria alignment of text-to-image models in urban planning. It asks:

How do different communities actually want public spaces to look and feel—and can models reflect that?

Across a two-year partnership with 30 community organizations in Montréal, we gathered local concepts of inclusive design, then turned them into structured preferences that can steer generative models toward context-aware outputs.

Why This Matters

Global, single-objective alignment often flattens local norms. Pluralistic alignment keeps multiple, sometimes conflicting, values in play—like Accessibility, Safety, Comfort, Invitingness, Inclusivity, and Diversity—so visualizations can reflect who a place is for as much as what it contains. That’s crucial when designs affect marginalized users differently.

How We Did It

Participants & Process: Two years with 30 orgs; workshops, interviews, and participatory prompt design.
From concepts to criteria: 634 community-defined ideas distilled into six criteria.
Images & Annotations: 13,462 images, 37,710 pairwise comparisons (multi-criteria sliders).
Modeling: Fine-tuned Stable Diffusion XL with Direct Preference Optimization (DPO) using multi-criteria signals.
Evaluation: In 2,200 new comparisons, annotators preferred the DPO-aligned model more often where annotation volume was higher; about half of judgments were neutral—evidence that values are heterogeneous and sometimes visually subtle.

What We Found

• No single “best” image: Improvements varied by criterion; Comfort/Invitingness benefited most from data volume; Inclusivity/Diversity remained harder to judge and to render consistently.
• Identity matters: Preference patterns differed across participant identities—supporting intersectional data over an “average user.”
• Prompts matter, too: Human-authored prompts produced more distinctive outputs—and more decisive annotations—than LLM-generated ones.
• Neutral is a signal: High neutrality often meant subtle trade-offs or ambiguous visuals, not noise—pointing to richer objective formulations beyond single binary labels.

Where This Can Be Used

• Public engagement: Generate options that mirror local priorities and surface trade-offs in workshops.
• Design iteration: Rapidly visualize accessibility, comfort, or safety tweaks grounded in multi-criteria feedback.
• Policy & research: Compare criteria across groups; explore steerable or personalized alignment strategies for different communities.

Visuals

A two-year co-creation process: outreach → criteria → prompts → images → multi-criteria comparisons → model tuning.

---

Distillation pipeline: Accessibility · Safety · Comfort · Invitingness · Inclusivity · Diversity.

---

Persistent gaps: intersectional identities consistently differ in their evaluations compared to majority groups.

---

More: arXiv · OpenReview / ICML 2025 · Project & Dataset (mid-space.one)

Tags: Pluralistic Alignment · Text-to-Image · Inclusive Urbanism · Multi-Criteria Preferences · Accessibility · Montréal