Reward Simulator

How it works

1. Select or upload an image that represents AI-generated content
2. The system finds similar images that might have influenced the generation in a database of 10M images (Open Images)
3. Based on your parameters, it calculates potential rewards for:
   • Original image authors
   • Rights owners (e.g., stock photo companies, galleries)

Warning

1. This demonstrator calculates similarities using the Open Images database alone. However, the image generators do not specify whether they have been trained with datasets from this database. As a result, calculations can be more or less biased depending on the generator used.
2. Furthermore, the infrastructures used are designed for limited use. More generally, this demonstrator is proposed for teaching and academic purposes. Its designers decline all responsibility for the relevance of the results, as well as for the user experience, which are provided without guarantee.

Key assumptions

• Attribution scores indicate the level of influence of training images
• Rewards are distributed based on subscription revenue
• Calculations use simplified models and are for demonstration purposes

Use cases

• Explore fair compensation models for AI training data
• Simulate different revenue sharing scenarios
• Understand the relationship between model training and attribution

Technical details

Image similarity search powered by DINOv2 and FAISS. DINOv2 is used to extract image features of all the images in the database (10M images from Open Images). FAISS is used to perform efficient similarity search to find the most similar images to the input image, and output the attribution scores.

• DINOv2 - Vision transformer for image feature extraction
• FAISS - Efficient similarity search library, more specifically:
  • HNSW (Hierarchical Navigable Small World) for approximate nearest neighbor search
  • PQ (Product Quantization) for memory-efficient vector compression