FaceAesthetic-HumanAI

A Research-Grade Dataset for Studying Human and AI Evaluation of Perceived Facial Attractiveness

---

Dataset Description

FaceAesthetic-HumanAI is a research-grade dataset for studying how humans and AI models evaluate perceived facial attractiveness. The dataset contains facial photographs with structured metadata, human attractiveness ratings from multiple annotators, and AI model predictions from multiple vision-language models.

Key principle: This dataset studies evaluation behavior, not objective beauty. All ratings reflect subjective perceptions of evaluators.

Version: 1.0.0 (Design Specification — data collection forthcoming)

---

Dataset Summary

This repository contains the complete design specification, schema, annotation guidelines, evaluation prompts, and reference code for the FaceAesthetic-HumanAI dataset. The dataset is designed to investigate how humans and AI models evaluate perceived facial attractiveness, and which visual factors contribute most strongly to these evaluations.

Current status: Design phase. Data collection pipeline is documented but no real human data has been collected yet. Synthetic demonstration records are provided for schema validation only.

---

Supported Tasks

• Human-AI agreement analysis
• Facial aesthetics perception research
• AI bias detection in visual evaluation
• Inter-rater reliability studies
• Model benchmarking for aesthetic prediction
• Presentation effect analysis (lighting, pose, expression)

---

Languages

• Image annotations: English
• Metadata: English
• Dataset documentation: English

---

Dataset Structure

Data Instances

Each instance contains:

• image: Facial photograph (JPEG)
• image_id: Unique identifier (FACE_000001)
• subject_id: Subject identifier (SUBJECT_000001)
• human_rating_mean: Mean human attractiveness rating (0-10)
• human_rating_std: Standard deviation of human ratings
• human_rating_count: Number of human annotators
• ai_predictions: Dict of AI model predictions
• image_characteristics: Pose, expression, lighting, etc.
• facial_attribute_ratings: Symmetry, proportion, etc.

Data Fields

Field	Type	Description
image	Image	Facial photograph (JPEG)
image_id	string	Unique identifier (FACE_000001)
subject_id	string	Subject identifier (SUBJECT_000001)
gender_presentation	string	male/female/unknown/non-binary
age_group	string	18-24/25-34/35-44/45-54/55+
image_type	string	portrait/selfie/studio/outdoor/candid/ID-photo
pose	string	front/three-quarter/profile/lookup
expression	string	neutral/smile/other
lighting	string	natural/studio/low-light/mixed/backlit
human_rating_mean	float	Mean human attractiveness rating (0-10)
human_rating_median	float	Median human rating
human_rating_std	float	Standard deviation of human ratings
human_rating_count	int	Number of human annotators
human_rating_confidence	float	Confidence in human rating
annotator_agreement_icc	float	ICC inter-rater reliability
ai_predictions	dict	AI model predictions (model_name → prediction)
human_ai_agreement	dict	Human-AI agreement metrics
image_characteristics	dict	Detailed image metadata
facial_attribute_ratings	dict	Symmetry, proportion, etc.
experiment_metadata	dict	Same-person, transformation, hard-negative info
provenance	dict	Source, consent, license, ethical review

Data Splits

Split	Images	Subjects	Purpose
Train	TBD	TBD	Model training
Validation	TBD	TBD	Hyperparameter tuning
Test	TBD	TBD	Final evaluation

Important: Splits will be subject-level — no subject will appear in multiple splits.

---

Dataset Creation

Curation Rationale

Existing facial aesthetics datasets suffer from:

1. Single-rater or few-rater designs
2. No AI model evaluation component
3. Limited metadata on image characteristics
4. No same-person multiple-image variants
5. No controlled transformation experiments

FaceAesthetic-HumanAI addresses these gaps.

Source Data

• Consented participants: Primary source (with IRB approval)
• Licensed datasets: Secondary source (with attribution)
• Synthetic images: For demonstration and augmentation

Data Collection

1. Image capture/collection with informed consent
2. Human annotation (20+ annotators per image)
3. AI model evaluation (multiple models, standardized prompts)
4. Quality control and validation
5. Metadata compilation

Annotation Annotators

• 100+ annotators from diverse backgrounds
• 20+ ratings per image
• Quality control via attention checks and gold-standard images
• Annotators may skip any image

---

Uses

Intended Uses

• Research on human-AI agreement in visual evaluation
• Benchmarking AI models for aesthetic prediction
• Studying bias in facial attractiveness evaluation
• Analyzing presentation effects on perceived attractiveness
• Training models to predict human aesthetic preferences

Out-of-Scope Uses

This dataset must NOT be used to:

• Rank individuals or groups by "attractiveness"
• Create "beauty scoring" applications
• Make hiring, admissions, or social decisions
• Develop surveillance or identification systems
• Reinforce harmful beauty standards
• Discriminate against any person or group

---

Bias, Risks, and Limitations

Known Biases

• Annotator bias: All annotators bring personal/cultural preferences
• AI bias: Models may reflect training data biases
• Presentation bias: Lighting, pose, and expression influence ratings
• Demographic bias: Dataset composition may not represent all populations

Limitations

• Subjective task with inherent disagreement
• Cultural diversity limited in v1.0
• AI predictions depend on prompt wording
• Same-person experiment limited to 50 subjects
• Cross-sectional design (no longitudinal data)

---

Ethics

• All real human images obtained with informed consent
• IRB approval obtained
• Participants have right to withdraw
• Dataset does NOT establish objective attractiveness
• Ratings reflect subjective perceptions only
• Not intended to rank individuals or groups

---

Research Questions

Primary: How do humans and AI models evaluate perceived facial attractiveness, and which visual factors contribute most strongly to these evaluations?

Secondary:

1. Do humans agree with each other?
2. Do different AI models agree with each other?
3. Do AI models agree with humans?
4. Which facial characteristics correlate with perceived attractiveness?
5. How much do lighting, pose, expression, grooming, and image quality influence ratings?
6. Does the same person's attractiveness score change significantly across different photographs?
7. Can AI predict the average human rating?
8. Where do AI and human judgments disagree?
9. Are some visual attributes disproportionately influential?
10. How stable are attractiveness predictions under controlled image transformations?

---

AI Model Evaluation

Evaluated Models

Model	Type	Source
GPT-4o	VLM	OpenAI
Claude 3.5 Sonnet	VLM	Anthropic
Gemini 1.5 Pro	VLM	Google
LLaVA-1.6	VLM	Open
Qwen-VL-Max	VLM	Alibaba
InternVL-2	VLM	Open

Evaluation Metrics

Metric	Description
Spearman ρ	Rank correlation with human ratings
Pearson r	Linear correlation with human ratings
MAE	Mean absolute error
RMSE	Root mean square error
ICC	Intraclass correlation coefficient
ECE	Expected calibration error

---

Experiments

1. Same-Person Multiple-Image Experiment

Multiple photographs of the same consenting subject under different conditions (lighting, pose, expression). Measures within-subject variance.

2. Controlled Transformation Experiment

Same image transformed by changing only one variable (lighting, crop, sharpness, etc.). Isolates presentation effects.

3. Hard Negatives

Examples where simple heuristics fail:

• High symmetry but moderate rating
• Lower symmetry but high rating
• Excellent image quality but moderate rating
• Poor lighting but strong facial aesthetics

---

Usage

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("IsmailTasdelen/FaceAesthetic-HumanAI")

train_dataset = dataset["train"]
val_dataset = dataset["validation"]
test_dataset = dataset["test"]

Basic Analysis

import pandas as pd

df = train_dataset.to_pandas()

# Rating distribution
print(f"Mean rating: {df['human_rating_mean'].mean():.2f}")
print(f"Rating std: {df['human_rating_std'].mean():.2f}")

# Filter by characteristic
studio_images = df[df['image_characteristics'].apply(lambda x: x['lighting'] == 'studio')]
print(f"Studio mean: {studio_images['human_rating_mean'].mean():.2f}")

Model Evaluation

from scipy import stats
from sklearn.metrics import mean_absolute_error

def evaluate_model(human_ratings, model_predictions):
    spearman_corr, _ = stats.spearmanr(human_ratings, model_predictions)
    mae = mean_absolute_error(human_ratings, model_predictions)
    return {"spearman": spearman_corr, "mae": mae}

---

Citation

@dataset{faceaesthetic_humanai_2025,
  title={FaceAesthetic-HumanAI: A Research-Grade Dataset for Studying Human and AI Evaluation of Perceived Facial Attractiveness},
  author={Tasdelen, Ismail and [Co-authors]},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/IsmailTasdelen/FaceAesthetic-HumanAI},
  license={CC-BY-NC-4.0},
  version={1.0.0}
}

---

License

This dataset is licensed under CC BY-NC 4.0.

---

Maintenance

• Version 1.0: Initial release (design specification)
• Planned v1.1: Expanded cultural diversity
• Planned v2.0: Longitudinal component, video support
• Issue tracker: GitHub repository

---

Contact

For questions or issues, please open an issue on the GitHub repository or contact the dataset maintainers.

---

Last updated: 2025-06-18