shaghayegh/stylistic-bias-dataset

Link to the paper

https://arxiv.org/abs/2606.20527

Dataset Description

StylisticBias is a controlled benchmark for evaluating fine-grained, attribute-level social bias in multimodal large language models (MLLMs). It addresses a key limitation of prior work: most bias studies compare different individuals or demographic groups, making it impossible to disentangle appearance-driven effects from identity-level variation.

StylisticBias keeps identity fixed and varies one visual attribute at a time, enabling direct measurement of how specific visual cues — such as hair style, eyewear, tattoos, or fashion style — shift an MLLM's social judgments about a person.

• Paper: StylisticBias: A Few Human Visual Cues Drive Most Social Biases in MLLMs
• Authors: Shaghayegh Kolli*, Timo Cavelius* (Equal Contribution), Nafiseh Nikeghbal, Samantha Dalal, Jana Diesner
• Code: github.com/timo-cavelius/StylisticBias
• License: CC-BY 4.0
• Languages: English (evaluation prompts and scenario labels)
• Modalities: Image + Text

---

Dataset Summary

Property	Value
Total images	~25,000
Base faces	500 synthetic photorealistic identities
Variations per face	~50 single-attribute edits
Demographic configurations	90 (3 age × 2 gender × 5 ethnicity × 3 body type)
Attribute categories	12 (skin, hair color, hair length, hair style, facial hair, makeup, lip makeup, eyewear, piercings, tattoos, headwear/accessories, fashion style)
Individual attribute values	~55 (full grid); 34 used in evaluation after reduction
Split	train (full dataset)
Image resolution	1024×1024 px
Format	Parquet (with embedded images)

---

Dataset Structure

Columns

Column	Type	Description
id	string	Unique identifier for the variation record
face_group	string	ID of the base face this variation belongs to
asset_kind	string	Type of image: face (head-and-shoulders portrait), body (full-body, for fashion variations)
image	image	The generated variation image
image_file	string	Filename of the variation image
image_relative_path	string	Relative path to the variation image within its face group folder
metadata_relative_path	string	Relative path to the JSON metadata file for this variation
age	string	Age group of the base face: young adult, middle-aged adult, elderly
gender	string	Perceived gender of the base face: male, female
ethnicity	string	Ethnicity label of the base face: Asian, African, European, Middle Eastern, Latino
body_type	string	Body type of the base face: thin, normal, obese
prompt	string	The exact generation prompt used to produce this variation
timestamp	string	ISO 8601 timestamp of image generation
base_image_name	string	Filename of the base (reference) face image
characteristics_json	string	JSON string with the base face's demographic attributes
variation_json	string	JSON string with the single attribute varied in this image (e.g., {"hair_style": "Messy"})
metadata_json	string	Full JSON metadata record including characteristics, variation, and generation details

Asset Kinds

• face: Head-and-shoulders portrait. Used for all attribute categories except fashion (skin irregularities, hair, makeup, eyewear, piercings, tattoos, accessories).
• body: Full-body portrait. Used exclusively for fashion style variations, which require a full-body view to be meaningful.

Attribute Space

The full variation space is as follows. Values excluded from the evaluation subset (but present in the dataset) are marked with *.

Attribute	Values
Skin irregularities	Freckles, Acne, Scars*, Moles
Hair color	Black, Brown, Blonde, Red*, Gray, Unnatural color
Hair length	Bald, Short, Medium*, Long
Hair style	Messy, Slicked back, Ponytail*, Braid*, Bun*, Afro*, Buzz cut*, Mohawk
Facial hair (M only)	Clean-shaven, Stubble*, Mustache*, Full beard
Eyewear	Thick-rimmed, Thin metal*, Sunglasses
Makeup (F only)	Light, Heavy
Lip makeup (F only)	Neutral*, Red lipstick, Bold*
Piercings	Single nose, Single lip*, Single eyebrow*, Multiple, Earrings*
Tattoos	Facial tattoo
Accessories	Cap, Beanie, Hat*, Headscarf*
Fashion style	Professional/Business formal, Formal/Evening wear, Casual, Smart casual, Sporty/Athletic wear, Streetwear, Functional/outdoor wear, Luxury/High fashion*, Vintage/Retro, Worn/Distressed clothing, Daring/Provocative*

*Excluded from the evaluation subset for plausibility or redundancy reasons (see paper Appendix C.1).

---

Dataset Creation

Source Data

All images are fully synthetic — no real photographs of individuals were used. Base faces were generated using Imagen 4 (Google DeepMind, 2025). Attribute variations were produced using Nano Banana (Gemini 2.5 Flash Image) (Comanici et al., 2025).

Base Face Generation

500 base faces were generated from a structured demographic prompt family. The Cartesian product of four demographic attributes — age (young adult, middle-aged, elderly) × gender (male, female) × ethnicity (Asian, African, European, Middle Eastern, Latino) × body type (thin, normal, obese) — yields 90 unique configurations, from which 500 identities were sampled for broad demographic coverage.

Base prompts enforce: front-facing pose, neutral expression, head-and-shoulders framing, plain white background, soft even lighting, no accessories or makeup, natural skin texture.

Observed base-face distribution:

• Gender: 274 male, 226 female
• Body type: 186 normal, 160 obese, 154 thin
• Ethnicity: 110 Asian, 109 African, 101 European, 95 Middle Eastern, 85 Latino
• Age: 260 young adult, 124 middle-aged, 116 elderly

Variation Generation

For each base face, single-attribute variations were generated by modifying exactly one feature at a time while explicitly preserving the base identity. Fashion-style variations use a full-body prompt template; all other variations use a head-and-shoulders portrait template.

Human Validation

90% of all generated images (base faces and variations) were manually reviewed, checking for: demographic plausibility, identity consistency across variations, and correct realization of the intended attribute change without unintended artifacts. 98% of reviewed images passed validation; failures were regenerated and re-reviewed before inclusion.

---

Intended Use

StylisticBias is intended as a benchmark for bias evaluation in multimodal large language models. It is designed for researchers studying:

• Attribute-level visual bias in vision-language models
• How specific appearance cues shift model social judgments
• Intersectional effects of demographic context and visual attributes
• Fairness auditing of deployed multimodal AI systems

Out-of-Scope Use

This dataset is not intended for:

• Re-identification of real individuals (all faces are synthetic)
• Any application that would operationalize the social judgment scenarios evaluated here (e.g., automated hiring or judicial support systems)

---

Evaluation Setup

In the companion paper, six open-source MLLMs were evaluated on this benchmark:

Model	Provider	Params
LLaVA-v1.6-Mistral-7B	LLaVA Team	7B
Qwen3-VL-8B-Instruct	Alibaba	8B
Pixtral-12B	Mistral AI	12B
InternVL3-14B	OpenGVLab	14B
Gemma-3-12B-IT	Google DeepMind	12B
Gemma-4-E4B-IT	Google DeepMind	4B (effective)

Models were queried on 25 binary social judgment scenarios spanning four categories: Personality & Social, Interpersonal, Behavioral, and Socioeconomic & Appearance. Each (image, scenario) pair was evaluated under 4 prompt orderings × 3 random seeds = 12 prompts, for approximately 4.72 million judgment calls per model and ~28.3 million total.

Limitations

1. Synthetic images only. The benchmark uses controlled synthetic data rather than real photographs. This is a deliberate design choice to enable single-attribute controlled variation, but the benchmark may not capture the full distribution of real-world photographs. Conclusions characterize model behavior in a controlled visual setting.

2. Curated demographic and attribute scope. The study covers a focused set of demographic groups and visual attributes. Many other identities and appearance dimensions exist and are not represented. The broad demographic categories used (e.g., ethnicity labels) are social constructs and may not reflect the full complexity of real-world identity.

3. Input-level effects only. The benchmark measures how visual attributes shift model outputs; it does not explain the underlying mechanisms producing these effects.

---

Ethical Statement

This dataset studies how visual attributes drive social judgments in MLLMs deployed in consequential settings. All faces are fully synthetic and do not represent or resemble any real individual. Synthetic face generation reduces privacy risks but may reproduce stereotypical associations from the generative models' training data.

We acknowledge dual-use risks: the methodology used to construct this benchmark could in principle inform adversarial appearance manipulation in automated pipelines. We release StylisticBias specifically to support fairness auditing and bias attribution research.

The demographic categories and social judgment scenarios used in this benchmark are social constructs that can reflect stereotypical perceptions and normative expectations. They are operationalized here strictly for the purpose of measuring and quantifying bias in model behavior, not to endorse or reproduce those judgments.

---

Citation

If you use this dataset, please cite:

@misc{stylisticbias2026,
  title  = {StylisticBias: A Few Human Visual Cues Drive Most Social Biases in MLLMs},
  author = {Shaghayegh Kolli and Timo Cavelius and Nafiseh Nikeghbal and Samantha Dalal and Jana Diesner},
  year   = {2026},
}

---

Dataset Card Authors

Shaghayegh Kolli*, Timo Cavelius* (Equal Contributaion), Nafiseh Nikeghbal, Samantha Dalal, Jana Diesner