OpenPeople

Dataset of synthetic people for testing generative image/video pipelines at scale.

What is OpenPeople?

OpenPeople provides a curated set of synthetic "people" designed for testing generative image and video pipelines. All people in this library are entirely synthetic. They do not represent real individuals (any similarity is purely coincidental).

Why use synthetic people?

• Safe testing: Test your pipelines without using real people's images
• Consistent subjects: Stable, reproducible characters across runs
• Prompt provenance: Every image includes the prompts used to generate it
• Diversity by design: Curated dataset covers various demographics

How was this dataset created?

Each synthetic person is generated through a multi-stage pipeline that ensures visual consistency across all assets.

1. Character Definition

Each person starts with a structured metadata.json defining their characteristics. These are randomly generated to ensure diversity:

Demographics are randomized across:

• Age ranges: 18-25, 25-35, 35-45, 45-55, 55-65, 65+
• Sex: male, female
• Ethnicity: Mixed heritage with weighted percentages (e.g., "14% Western Asian, 50% European, 36% African")
• Skin tone: Fitzpatrick scale I–VI

Visible traits include randomized:

• Hair: texture, length, color, style (e.g., "Very long, thick, naturally wavy curls")
• Body type: slim, athletic, average, curvy, large
• Height: short, average, tall

Additional traits add unique details like jawline, eyebrows, hairline, and other distinguishing features.

{
  "person_id": "P050",
  "character_details": {
    "demographics": {
      "age_range": "18-25",
      "sex": "male",
      "ethnicity": "14% Western Asian, 50% European, 36% African",
      "skin_tone": "Fitzpatrick III"
    },
    "visible_traits": {
      "hair": { "description": "Very long, thick, naturally wavy curls cascading over the shoulders." },
      "body_type": "large",
      "height": "average"
    },
    "additional_traits": {
      "hairline": "Center-parted hair with a neat hairline",
      "hands": "Long, slender fingers resting gently on the chin."
    }
  }
}

2. Multi-Stage Image Generation

Images are generated in a specific order, where each stage uses previous outputs as reference to maintain identity consistency:

┌─────────────────┐
│ Character JSON  │
└────────┬────────┘
         ▼
┌─────────────────┐
│    Preview      │  Temp full-body preview
└────────┬────────┘
         ▼
┌─────────────────┐
│ Studio Portrait │  Professional headshot
└────────┬────────┘
         ▼
┌─────────────────┐
│ Studio Posture  │──────────────┐  Full-body studio shot
└────────┬────────┘              |
         ▼                       ▼
┌─────────────────┐     ┌─────────────────┐
│ Character Sheet │     │ Emotions Sheet  │  Reference sheets
└────────┬────────┘     └────────┬────────┘
         │                       │
         └───────────┬───────────┘
                     ▼
         ┌─────────────────────┐
         │   Amateur Photos    │  Casual photos
         └─────────────────────┘

3. Reference Chaining

The key to consistency is reference chaining — each generation stage receives images from previous stages:

• Studio Portrait: Uses the preview image to establish facial features
• Studio Posture: Uses the portrait to maintain face consistency in full-body shot
• Character Sheet: Combines portrait + posture with a layout template
• Emotions Sheet: Uses portrait with an emotions grid template
• Amateur Photos: Use the character/emotions sheets as reference for varied contexts

4. Generation Model

All images in the curated dataset are generated using Gemini 3 Pro image generation with carefully crafted prompts for each asset type. The prompts emphasize analytical photography, visible imperfections, and consistent styling.

Installation

pip install openpeople

Quick Start

import openpeople

# List all curated people
people = openpeople.curated.list()
print(f"Found {len(people)} synthetic people")

# Get a specific person
person = openpeople.curated.get("P001")

# Get a random person (non-deterministic)
random_person = openpeople.curated.random()

# Get a random person (deterministic with seed)
seeded_person = openpeople.curated.random(seed=42)

# Sample multiple people without replacement
sample = openpeople.curated.sample(n=3, seed=1234)

Working with Assets

Each person comes with multiple image assets:

person = openpeople.curated.get("P001")

# Get path to an asset
path = person.asset_path("studio_selfie")
print(path)

# Load image directly (requires openpeople[images])
image = person.load_image("studio_selfie")
image.show()

⚠️ Disclaimer

Please verify on your own before usage.

These synthetic people are provided for testing and development purposes only. OpenPeople contains:

• No names or personal identifiers
• No real individuals
• No data suitable for identification purposes

Users are responsible for ensuring their use complies with applicable laws and ethical guidelines.

License

MIT License - see LICENSE for details.

Contributing

Contributions are welcome! Please ensure:

1. All synthetic people remain entirely fictional
2. No names or real-person references
3. Used prompts must be documented
4. Tests pass before submitting PRs

Built with 💜 by Prompt Haus