mengtingwei/emotion_bias

Emotion Bias in Synthetic Face Generation

Description

This dataset accompanies the paper "Happy Young Women, Grumpy Old Men? Emotion Prompts as Demographic Selectors in AI Image Generation".

It contains 56,000 synthetic face images generated by eight state-of-the-art text-to-image (T2I) models across seven emotion prompt conditions, along with demographic attribute annotations (gender, race, age) and perceived attractiveness labels for each image.

The dataset is designed to support research on demographic bias, intersectional fairness, and emotion-conditioned generation in T2I models.

Models Audited

Model	Developer	Architecture	Region
FLUX-Schnell	Black Forest Labs	DiT	Western
Proteus V0.3	HuggingFace community	UNet	Western
SD3.5 medium	Stability AI	DiT	Western
SANA 1.5	NVIDIA	DiT	Western
Hunyuan-DiT	Tencent	DiT	Chinese
Qwen-Image	Alibaba	DiT + LoRA	Chinese
Kolors	Kuaishou	UNet	Chinese
Wan2.1	Wan-AI	Video DiT	Chinese

Emotion Conditions

Seven prompt conditions were used:

• Neutral (baseline): "A photorealistic portrait of a person, front-facing"
• Six emotions: Happy, Sad, Angry, Surprised, Disgusted, Fearful
  • Template: "A photorealistic portrait of a [emotion] person, front-facing"

Dataset Statistics

• Total images: 56,000 (8 models × 7 conditions × 1,000 images)
• Resolution: 1024×1024 (Wan2.1: 832×480)
• Demographic annotations: gender (male/female), race (White/Black/Asian/Others), age (0–9, 10–19, 20–39, 40–59, 60+)
• Attractiveness annotations: low / medium / high (three-class)

Annotation Tools

• Gender, Race, Age: FairFace (Karkkainen & Joo, WACV 2021)
• Attractiveness: ResNet-based classifier trained on the AHEAD dataset (Gulati 2025)

Key Findings

• All eight models strongly overrepresent young (20–39) faces (>80% in six models)
• Most models overrepresent White-coded faces relative to global population baselines
• Emotion prompts act as demographic selectors: negatively valenced emotions (Anger, Disgust, Fear, Sadness) consistently shift outputs toward male-coded, White-coded, and older faces
• The young × female × Black intersectional combination appears at only 1.3% on average (suppression factor 0.25× relative to population expectation)
• Western and Chinese models exhibit broadly similar bias patterns despite different development contexts

Data Collection

Images were generated using default inference hyperparameters for each model. Images in which no face was detected by the dlib-based detector (bundled with FairFace) were discarded and replaced with newly sampled generations, ensuring exactly 1,000 valid face images per model–condition cell.

License

This dataset is released under CC BY 4.0. It is free to use for both research and commercial purposes with attribution.

Related Resources

• Code: https://github.com/weimengting/EmotionBias-SyntheticFaces