Research map / Image generation / Text-to-image diffusion
GeoDiv: Framework for Measuring Geographical Diversity in Text-to-Image Models
Abhipsa Basu, Mohana Singh, Shashank Agnihotri, Margret Keuper, Venkatesh Babu Radhakrishnan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Gemini-2.5-flash instead of weaker VLMs as the VQA judge (ablation) | GeoDiv VDI · VQA validation, overall | accuracy: 0.68 (LLaVA) -> 0.86 (Gemini-2.5-flash) | Improved |
| compare VLM judges on SEVI Maintenance alignment (Gemini vs gpt-4o) (ablation) | GeoDiv SEVI · SEVI validation, Maintenance | Spearman correlation: Gemini 0.69 vs gpt-4o 0.76 | Mixed |
| newer diffusion model versions (SD2.1 to FLUX.1) versus older | T2I models · GeoDiv overall score | GeoDiv score: 0.4251 (SD2.1) -> 0.3153 (FLUX.1) | Hurt |
| FLUX.1 polished generation versus SD models | FLUX.1-dev · Mean Maintenance rating | rating: Maintenance 3.23 (SD2.1) -> 4.73 (FLUX.1); diversity 0.15 | Mixed |
| real-world GeoDE images versus synthetic images | image dataset · Entity-Appearance Diversity | Diversity-Score: 0.44 -> 0.60 | Improved |
| real-world GeoDE images versus synthetic images | image dataset · Background-Appearance Diversity | Diversity-Score: 0.31 -> 0.42 | Improved |
| prompt for wealthy countries (Japan, UAE, UK) versus India, Nigeria, Colombia | T2I models · Mean Affluence rating | rating: 2.31 -> 3.53 | Improved |
| use Vendi-Score instead of GeoDiv for non-entity axes | Vendi-Score · Correlation with GeoDiv Maintenance | Pearson correlation: 0.56 (entity) vs 0.06 (maintenance) | Hurt |
| use open-source Qwen2.5-VL in place of Gemini for GeoDiv (ablation) | GeoDiv · agreement across four diversity axes | average correlation: rho = 0.83 | No clear effect |
| country wealth (GDP) versus GeoDiv score | T2I models · GeoDiv score by country | correlation: rho = -0.27 (GDP), -0.28 (per capita) | Hurt |
Closest papers by meaning
- No Filter: Cultural and Socioeconomic Diversity in Contrastive Vision-Language Models (NeurIPS 2024 · similarity 0.90)
- OASIS Uncovers: High-Quality T2I Models, Same Old Stereotypes (ICLR 2025 · similarity 0.88)
- Multilingual Diversity Improves Vision-Language Representations (NeurIPS 2024 · similarity 0.88)
- Beyond Aesthetics: Cultural Competence in Text-to-Image Models (NeurIPS 2024 · similarity 0.87)
- GeoReward: Mitigating Contextual Variable Overestimation in Vision-Language Models for Cross-Market Preference Prediction (ICML 2026 · similarity 0.85)
- TerraVis: Towards Evaluation of World-Grounded Visual Consistency in Text-to-Image Generation via MLLM Workflows (NeurIPS 2026 · similarity 0.85)
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.
