Research map / Image generation / Text-to-image diffusion
LIVS: A Pluralistic Alignment Dataset for Inclusive Public Spaces
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces LIVS, a participatory multi-criteria pairwise preference dataset for inclusive Montreal public spaces, and uses it to DPO-fine-tune SDXL with four case studies showing modest, often neutral, gains.
Problem
Existing T2I alignment relies on global, single-objective preference data and crowdwork, which fails to capture intersectional, locally grounded criteria for inclusive urban public space design.
Approach
A two-year participatory process with 30 community organizations produced six locally defined criteria (Accessibility, Safety, Comfort, Invitingness, Inclusivity, Diversity) and slider-based pairwise annotations on SDXL images, used to DPO-fine-tune SDXL with majority-voted binary labels.
Key results
- Of 2,200 new comparisons, 700 (32%) favored the DPO-tuned SDXL, 300 (14%) the baseline, and 1,100 (50%) were neutral.
- Criteria with more training annotations (Comfort, Invitingness) showed stronger DPO gains; Inclusivity and Diversity had higher neutral rates.
- Human-authored prompts produced fewer neutral ratings than GPT-4o-generated prompts (Case Study III).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune Stable Diffusion XL with multi-criteria LIVS preferences using DPO | Stable Diffusion XL · LIVS evaluation set | preference rate: 32% vs. 14% | Helped |
| use multi-criteria DPO instead of baseline SDXL | Stable Diffusion XL · LIVS evaluation set | neutral rating rate: 1,100 (50%) | Mixed |
| use human-authored prompts instead of GPT-4o-generated prompts | SDXL and DPO-finetuned SDXL · LIVS prompt comparison | neutral rating rate | Helped |
| train with more preference annotations for a criterion | DPO-finetuned SDXL · LIVS evaluation set, Comfort and Invitingness | DPO preference | Helped |
| use multi-criteria DPO rather than single-criterion fine-tuning (ablation) | Stable Diffusion XL · LIVS public-space image examples | visual alignment with local preferences | Helped |
| evaluate images from different intersectional identity groups | SDXL outputs · LIVS, Accessibility and Safety | raw rating | Mixed |
Surprising findings
- Roughly half of the 2,200 final comparisons were neutral, so DPO gains were limited.
- Human-authored prompts yielded more decisive annotations than LLM-generated prompts.
Limitations
- Data come from one mid-sized Montreal city with relatively few participants.
- Multi-criteria slider labels were collapsed to binary via majority voting, discarding neutral and intra-sample disagreement.
- Test set of 2,200 annotations is modest relative to T2I complexity.
Experimental setup
Datasets: LIVS
Benchmarks: LIVS evaluation set (2,200 new annotations)
Metrics: preference share (DPO vs baseline vs neutral) neutral rate by criterion per-identity preference counts
Baselines: SDXL baseline
Builds on: Direct Preference Optimization Stable Diffusion XL
Closest papers by meaning
- UrbanFeel：A Comprehensive Benchmark for Temporal and Perceptual Understanding of City Scenes through Human Perspective (ICLR 2026 · similarity 0.85)
- SocialCounterfactuals: Probing and Mitigating Intersectional Social Biases in Vision-Language Models with Counterfactual Examples (CVPR 2024 · similarity 0.84)
- Cross-City Latent Space Alignment for Consistency Region Embedding (ICML 2025 · similarity 0.84)
- UrbanIng-V2X: A Large-Scale Multi-Vehicle, Multi-Infrastructure Dataset Across Multiple Intersections for Cooperative Perception (NeurIPS 2025 · similarity 0.84)
- Multi-Group Proportional Representations for Text-to-Image Models (CVPR 2025 · similarity 0.84)
- Whose View of Safety? A Deep DIVE Dataset for Pluralistic Alignment of Text-to-Image Models (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.