Technical Summary: Emotion-Driven Demographic Biases in Synthetic Face Generation

Problem Statement

While text-to-image (T2I) and multimodal large language models (MLLMs) have achieved high-fidelity synthetic face generation, the representational quality and bias of these systems remain poorly understood, particularly regarding the intersection of demographic attributes and emotional prompts. Prior research has largely focused on isolated demographic biases (e.g., gender or race alone) in neutral faces or on Western-trained models. This study identifies three critical gaps:

1. Intersectionality: Existing work often overlooks how multiple identities (gender, race, age) compound to create unique vulnerabilities and under-representations.
2. Emotion-Conditioned Bias: Most studies ignore how emotional prompts (e.g., "happy" vs. "angry") influence demographic distributions, despite evidence that facial expressions are tied to cultural norms and stereotypes.
3. Cross-Cultural Consistency: There is a lack of systematic comparison between models developed in Western contexts versus those developed in China, despite China's emergence as a major AI power with region-specific training data.

The central research questions address whether T2I models generate demographically imbalanced faces relative to global populations, how these biases differ between Western and Chinese models, and how emotional prompts shift demographic and attractiveness distributions.

Methodology

The authors conducted a systematic audit of eight state-of-the-art diffusion-based T2I models: four developed by Western organizations (FLUX-Schnell, Proteus V0.3, Stable Diffusion 3.5, SANA 1.5) and four by Chinese institutions (Hunyuan-DiT, Qwen-Image, Kolors, Wan2.1).

Experimental Design:

• Prompting: The models were prompted with identical English templates: "A photorealistic portrait of a [emotion] person, front-facing."
• Emotions: Six canonical emotions (happy, sad, angry, surprised, disgusted, fearful) based on Ekman's basic emotions framework were tested, alongside a neutral baseline.
• Scale: 1,000 images were generated per model per emotion, totaling 56,000 synthetic faces.
• Attribute Estimation: Generated faces were analyzed using state-of-the-art automated tools:
  • Demographics: The FairFace model was used to estimate gender, race, and age. Validation on benchmark datasets (CFD, FACES, APPA-REAL, FGNET) showed high accuracy (97% for gender, 90% for race, 73% for age).
  • Attractiveness: A dedicated model (based on ResNet and trained on the AHEAD dataset) estimated attractiveness levels (low, medium, high).
• Bias Metrics: The study employed information-theoretic metrics to quantify deviations:
  • Baseline Bias: Kullback-Leibler (KL), Jensen-Shannon (JS), and Total Variation Distance (TVD) divergences were calculated between model outputs and global population statistics (derived from UN data and country-to-race assignment strategies).
  • Intersectional Shift: KL and JS divergences measured shifts in the joint distribution of gender, race, and age.
  • Emotion-Conditioned Bias: Divergences were calculated between emotion-specific outputs and the neutral baseline to isolate the impact of emotional prompts.

Key Results

1. Persistent Demographic Bias (RQ1)
All eight models exhibited significant deviations from global population statistics, regardless of their origin.

• Gender: Most models showed pronounced gender imbalance. Four models (FLUX, Proteus, SD3, Qwen) over-represented males (60–90%), while two (SANA, Hunyuan) over-represented females (>80%). Only Qwen and Wan2.1 approached a balanced distribution, though none achieved parity.
• Race: Disparities were even more severe. Three models (Proteus, SANA, Kolors) generated over 80–90% White faces, despite White individuals comprising ~11% of the global population. Asian and Black populations were consistently under-represented.
• Age: A strong generative prior toward youth was observed. The 20–39 age group was heavily over-represented (>80% in six models), while children, adolescents, middle-aged adults, and especially the elderly (60+) were severely under-represented.
• Intersectionality: The joint distribution analysis revealed that specific combinations, such as "young × female × Black," had near-zero representation in most models, while "young × female × White" dominated.

2. Cross-Cultural Homogenization (RQ2)
Contrary to the expectation that Chinese models would reflect local demographics, the study found a surprising homogenization of biases.

• Similarity: Western and Chinese models exhibited comparable patterns of over-representing White and young individuals.
• Race Anomaly: Despite China's predominantly East Asian population, most Chinese models generated a majority of White faces, closely mirroring Western models. Hunyuan was a notable exception, generating ~90% Asian faces.
• Language Impact: Testing Chinese models with Chinese prompts (Appendix E) showed only moderate distributional shifts; the over-representation of White faces persisted, suggesting the bias stems from shared training data or architectural priors rather than prompt language alone.

3. Emotion-Conditioned Bias (RQ3)
Emotional prompts systematically shifted demographic distributions, exacerbating existing biases.

• Negative Emotions: Prompts for negative, high-arousal emotions (Anger, Disgust, Fear) induced the largest demographic shifts. These prompts consistently increased the probability of generating older, male, and White faces while decreasing the representation of young, female, and Asian faces.
• Positive Emotions: "Happy" prompts resulted in the smallest deviations from the neutral baseline, often reinforcing the default bias toward young, attractive, and White faces.
• Attractiveness: Negative emotions correlated with a significant increase in low-attractiveness scores, while happiness slightly increased high-attractiveness scores, aligning with the "attractiveness halo effect."
• Model Sensitivity: Western models generally exhibited larger KL divergences (greater sensitivity to emotion) than Chinese models, though the direction of the shifts was consistent across both groups.

Significance and Claims

The paper claims to provide the first systematic, cross-cultural, emotion-conditioned audit of T2I-generated faces. Its primary contributions are:

1. Quantitative Framework: A method for measuring representational biases using information-theoretic metrics across intersecting attributes.
2. Evidence of Homogenization: The finding that models from distinct cultural ecosystems (Western and Chinese) converge on a singular, Western-centric "default" representation, challenging the notion that regional training data necessarily leads to culturally distinct outputs.
3. Emotion as a Bias Amplifier: The discovery that adding emotional cues does not merely change facial expression but fundamentally reshapes demographic distributions, often reinforcing ageism and racial/gender stereotypes (e.g., associating anger with older males).

The authors argue that these biases are not merely artifacts of data composition but reflect structural priors in the latent space of generative models. They conclude that without active auditing and intervention, these systems risk reinforcing exclusionary stereotypes, particularly when used for "emotionally charged" imagery. The study advocates for transparency in training data, granular user controls to decouple emotion from demographic defaults, and proactive regulation requiring regular bias audits before large-scale deployment.

Limitations: The authors acknowledge that their audit is limited to a specific set of models and a predefined set of emotions and attributes. They note that their reliance on automated classifiers, while scalable, may introduce classification errors, and their framework is diagnostic rather than prescriptive regarding mitigation strategies.