nati1221/gcfairbench-v2

GCFairBench-100 (v3.2) — DOPP-CRAFT-GC experiment images

Interactive gallery: https://nati1221-craft-gc-gallery.static.hf.space/
Search by prompt ID or text; compare all 5 methods side-by-side; browse all 2,500 images.

2,500 unique text-to-image outputs for the DOPP-CRAFT-GC MSc thesis / research evaluation.

Field	Value
Prompts	100 (GCFairBench-100: 40 SSA, 15 each other region)
Methods	Base SD, PromptAug, PromptAug-Explicit, FairImagen-GC, DOPP-CRAFT-GC v3.2
Seeds	42, 123, 456, 789, 1011
Total images	2,500
Model	Realistic Vision V5.1
Evaluation	CLIP ViT-B/32 (image_eval.csv)

Headline results (v3.2, 2,500 images, 5 seeds)

Method	CoFS	GP	SSA-BFS	GCFair
Base SD	0.914	1.000	+0.000	0.259
PromptAug	0.868	0.699	+0.140	0.470
PromptAug-Explicit	0.874	0.852	+0.464	0.926
FairImagen-GC	0.910	0.849	−0.092	0.000
DOPP-CRAFT-GC v3.2	0.851	0.965	+0.397	0.950

DOPP-CRAFT-GC achieves the best composite GCFair (0.950) and gender parity (0.965), with SSA Black-fidelity Δ of +0.397 (Wilcoxon p ≈ 4.9e-10, Cohen's d ≈ 1.38). FairImagen alone reduces SSA representation, motivating the geo-cultural anchor steering.

GCFairBench-100 design

Region	Prompts	Images
Sub-Saharan Africa	40	1000
South Asia	15	375
Southeast Asia	15	375
MENA	15	375
Latin America	15	375

12 prompts explicitly reference Ethiopia or Addis Ababa.

Files

• manifest.json — prompt ID, region, method, seed, path
• paper_results.json — metrics + statistical tests
• image_eval.csv — per-image CLIP race/gender probes
• metrics.csv — method summary
• pilot_images/{method}/*.png — all images

Files

• manifest.json — prompt ID, region, method, seed, relpath (2,500 entries)
• paper_results.json — v3.2 metrics + statistical tests + SFV
• image_eval.csv — per-image CLIP race/gender probes (2,500 rows)
• metrics.csv, variance.csv, sfv.csv, main_table.tex
• pilot_images/{base,prompt_aug,prompt_aug_explicit,fairimagen,craftgc}/*.png

Citation

@misc{craftgc2026gcfairbench,
  title={GCFairBench-100: Geo-Cultural Fairness Benchmark for Text-to-Image Diffusion (DOPP-CRAFT-GC v3.2)},
  author={Abayneh, Natnael and Alemu, Abiy},
  year={2026},
  howpublished={\\url{https://huggingface.co/datasets/nati1221/gcfairbench-v2}}
}