nati1221/craft-gc-human-eval-results

DOPP-CRAFT-GC v3.2 — Human Evaluation Results

Blinded 3-evaluator human study collected via the Space nati1221/craft-gc-human-eval.

Each evaluator rated 15 region-balanced prompts. For every prompt, 5 method images (A-E, blinded order) were rated on two 1-5 scales (Realism & Quality, Cultural & Fair Representation) and one "best image" choice.

Files

• ratings.jsonl — one JSON object per image rating (all evaluators).
• submissions/ — one JSON snapshot per submission.

Row schema (ratings.jsonl)

field	meaning
evaluator_id	E1 / E2 / E3
background	self-reported evaluator background
timestamp	UTC ISO timestamp
prompt_id, region, seed	benchmark prompt identity
label	blinded label shown to evaluator (A-E)
method	de-blinded method (base / prompt_aug / prompt_aug_explicit / fairimagen / craftgc)
image_url	source image on nati1221/gcfairbench-v2
realism_quality	1-5
cultural_fair	1-5
is_best	1 if chosen best image for that prompt
comment	optional free text

Analysis

Run scripts/analyze_human_eval.py to aggregate per-method means and best-vote counts across all evaluators.