Legal & Copyright Guidelines

Overview

This document outlines the legal considerations and best practices for using the Lumira project, particularly regarding copyright, training data, and generated content.

Last Updated: January 8, 2026

⚠️ Important Legal Disclaimer

• This project is for educational and personal use
• Users are responsible for ensuring their use complies with applicable laws
• Consult a lawyer for commercial applications
• Generated content may have complex copyright status

---

Training Data Guidelines

Copyright Considerations

Training AI models on copyrighted artwork carries legal risks. According to the U.S. Copyright Office (2025):

• Training to mimic specific artists makes "copying easy to prove"
• Using copyrighted works without permission may constitute infringement
• Fair use defense is complex and case-by-case

Safe Training Data Sources

✅ RECOMMENDED:

1. Public Domain Works

   • Pre-1928 artworks (copyright expired)
   • Government-created works
   • Explicitly public domain datasets

2. Creative Commons Licensed Works

   • CC0 (Public Domain Dedication)
   • CC-BY (Attribution required)
   • CC-BY-SA (Attribution + ShareAlike)
   • Verify license terms before use

3. Your Own Artwork

   • Original works you created
   • Works where you hold copyright

❌ AVOID:

• Copyrighted artwork without permission
• Works by living artists without explicit consent
• Dataset with unclear licensing
• Commercial stock photography

Recommended Training Datasets

safe_datasets:
  - WikiArt Public Domain collection (pre-1928 works)
  - Rijksmuseum Open Data
  - Metropolitan Museum Open Access
  - Smithsonian Open Access
  - Library of Congress Public Domain
  - Your own original artwork (20-50 images)

Documentation Requirements

For each training session, document:

training_metadata = {
    "dataset_sources": ["Rijksmuseum Open Access", "Personal Collection"],
    "licenses": ["CC0", "Self-owned"],
    "date_ranges": ["1600-1900", "2024-2026"],
    "artist_names": ["None - public domain only"],
    "verification_date": "2026-01-08",
    "total_images": 45
}

---

Generated Content Copyright

Copyright Status

AI-generated artwork has complex copyright status:

• U.S. Copyright Office: Works must have "human authorship" for copyright
• Purely AI-generated works may not be copyrightable
• Works with significant human creative input may qualify
• Laws vary by jurisdiction and are evolving

Best Practices

1. Attribution

   • Credit the base model (e.g., "Created with Stable Diffusion 1.5")
   • Mention inspiration sources if applicable
   • Be transparent about AI involvement

2. Licensing Your Outputs

   recommended_license: "CC-BY 4.0"
   # Allows others to use with attribution
   # Acknowledges uncertain copyright status
   

3. Commercial Use

   • Verify model license permits commercial use
   • Check if training data licenses restrict commercial outputs
   • Consider additional licensing for commercial projects

---

API Usage Compliance

Unsplash API

Terms of Service Requirements:

• Attribution required for used images
• Cannot use images for competing services
• Rate limits: 50 requests/hour (free tier)
• Downloaded images used only as inspiration, not redistributed

Proper Attribution:

attribution = f"Photo by {photographer_name} on Unsplash"
# Store in database with each source image

Pexels API

Terms of Service Requirements:

• Attribution appreciated but not required
• Cannot sell unmodified photos
• Cannot create competing image service
• Rate limits: 200 requests/hour (free tier)

---

Guardrails & Safety Measures

Preventing Copyright Infringement

Implement these guardrails in your generation pipeline:

# Prohibited prompts (living artists)
BLOCKED_ARTISTS = [
    # Add names of living artists
    # Check regularly and update
]

# Prompt filtering
def check_prompt_safety(prompt: str) -> bool:
    """Prevent prompts that could generate infringing content"""
    prompt_lower = prompt.lower()

    # Block specific artist names
    for artist in BLOCKED_ARTISTS:
        if artist.lower() in prompt_lower:
            return False

    # Block style mimicry phrases
    risky_phrases = [
        "in the style of",
        "as painted by",
        "like [artist name]",
    ]

    # Allow style descriptors, block specific attribution
    return True

Content Monitoring

# Log all generations for review
generation_log = {
    "prompt": prompt,
    "negative_prompt": negative_prompt,
    "source_url": inspiration_url,
    "timestamp": datetime.now(),
    "safety_check": "passed"
}

---

Model Licensing

Stable Diffusion Models

Check license before use:

• SD 1.5: CreativeML OpenRAIL-M License

  • Allows commercial use
  • Prohibits illegal/harmful content
  • Cannot use to compete directly with Stability AI

• SDXL: CreativeML OpenRAIL++-M License

  • Similar terms to SD 1.5
  • Additional restrictions on derivatives

LoRA Weights

• LoRA weights inherit base model license
• Additional restrictions if trained on restricted data
• Document training data licenses

---

Privacy Considerations

API Keys & Credentials

• Never commit API keys to version control
• Use environment variables or secrets management
• Rotate keys regularly
• Limit key permissions to necessary scopes

User Data

If adding feedback or social features:

• Collect only necessary data
• Provide privacy policy
• Allow data deletion
• Comply with GDPR/CCPA if applicable

---

Compliance Checklist

Before training or deploying:

• Verified all training data is public domain or properly licensed
• Documented data sources and licenses
• Implemented artist name blocking in prompts
• Added attribution for API-sourced images
• Reviewed and accepted base model license terms
• Set up proper secrets management
• Added copyright disclaimer to outputs
• Configured rate limiting for API calls

---

Reporting Issues

If you discover:

• Potentially infringing outputs
• Training data with unclear licensing
• License violations

Actions:

1. Stop using the affected component
2. Document the issue
3. Consult legal advice if needed
4. Update this document with learnings

---

Resources

Legal Information

• U.S. Copyright Office - AI Reports
• Creative Commons Licenses
• Stability AI License

Safe Data Sources

• Rijksmuseum API
• Met Museum Open Access
• Public Domain Review

Model Licenses

• Stable Diffusion License
• Hugging Face Model Cards

---

Updates

This document will be updated as:

• Laws and regulations evolve
• New court cases provide guidance
• Model licenses change
• Community best practices develop

Version History:

• v1.0 (2026-01-08): Initial version

---

This document provides general information and is not legal advice. Consult a qualified attorney for specific legal questions.