Generative AI by Getty Images: Addressing Concerns and Building Better Models

Event: ? · Duration: 29 min · ▶ Watch on YouTube

Abstract

Getty Images presents its approach to generative AI, highlighting the importance of addressing customer concerns regarding legal risks, authenticity, and ethical considerations. The company has developed its own generative AI model, powered by NVIDIA Picasso, which is trained exclusively on licensed data to ensure commercial safety, fair compensation for creators, and authentic, diverse outputs. The presentation demonstrates how high-quality, diverse training data leads to more realistic and inclusive AI-generated images, contrasting with the biases and brand infringements seen in models trained on web-scraped data. Ultimately, Getty Images advocates for continued investment in human creativity and licensed content to build robust and socially reflective AI models for the future.

Speakers

  • Andoria — Getty Images

Talks (1)

  • 00:00:00 — Andoria: Generative AI by Getty Images: Addressing Concerns and Building Better Models
    • This talk discusses how Getty Images is approaching generative AI, focusing on addressing customer concerns about legal risks, authenticity, and diversity by using licensed, high-quality training data and compensating creators, ultimately aiming for AI models that better reflect society.

Key Takeaways

  • Getty Images’ generative AI model is built on commercially safe, licensed training data, ensuring no brands, logos, or celebrities are generated, and creators are compensated.
  • High-quality, diverse, and current training data is crucial for generative AI models to accurately reflect society and avoid biases seen in models trained on uncurated web data.
  • Compensating photographers and creators for their work is essential to incentivize the creation of new content, which in turn provides the necessary training data for future AI models.
  • Generative AI is a powerful tool for visualizing highly creative or difficult-to-capture concepts, but traditional camera-shot imagery remains vital for depicting real people, emotions, and societal nuances.
  • The quality and ethical sourcing of training data directly impact the authenticity, diversity, and overall utility of AI-generated images, influencing how well these models can keep pace with evolving societal trends.

Methods / Models / Datasets Mentioned

  • NVIDIA Picasso
  • Diffusion Model
  • FID (5K) score
  • CLIP score
  • Edify model architecture

Topics

Generative AI · Getty Images · AI ethics · Licensed data · Creator compensation · Diversity in AI · Image generation · Visual trends · Model training · AI bias · VisualGPS

Notes

Open for commentary — connections to other work, critiques, follow-up reading.