Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Shapley Values-Powered Framework for Fair Reward Split in Content Produced by GenAI

Alex Glinsky, Alexey Sokolsky

TL;DR

The paper tackles the problem of fairly compensating data contributors (artists) whose works influence GenAI outputs. It introduces a Shapley Values–driven framework that models each generated image as a coalition of content and style, using CLIP embeddings to quantify contributions and a black-box, prompt-based approximation of the Shapley value $\phi_i(v)$. Through single- and multi-style experiments—including Dreambooth and ImageMixer scenarios—the authors demonstrate how rewards can be split between artists and the model, while addressing biases and recognizability filtering. The findings show practical feasibility for local computation, enabling stakeholders to seek recompense for data contributions and fostering a more sustainable collaboration between data providers and model developers. This approach has implications for data provenance, artist rights, and the economics of future AI systems.

Abstract

It is evident that, currently, generative models are surpassed in quality by human professionals. However, with the advancements in Artificial Intelligence, this gap will narrow, leading to scenarios where individuals who have dedicated years of their lives to mastering a skill become obsolete due to their high costs, which are inherently linked to the time they require to complete a task -- a task that AI could accomplish in minutes or seconds. To avoid future social upheavals, we must, even now, contemplate how to fairly assess the contributions of such individuals in training generative models and how to compensate them for the reduction or complete loss of their incomes. In this work, we propose a method to structure collaboration between model developers and data providers. To achieve this, we employ Shapley Values to quantify the contribution of artist(s) in an image generated by the Stable Diffusion-v1.5 model and to equitably allocate the reward among them.

Shapley Values-Powered Framework for Fair Reward Split in Content Produced by GenAI

TL;DR

The paper tackles the problem of fairly compensating data contributors (artists) whose works influence GenAI outputs. It introduces a Shapley Values–driven framework that models each generated image as a coalition of content and style, using CLIP embeddings to quantify contributions and a black-box, prompt-based approximation of the Shapley value . Through single- and multi-style experiments—including Dreambooth and ImageMixer scenarios—the authors demonstrate how rewards can be split between artists and the model, while addressing biases and recognizability filtering. The findings show practical feasibility for local computation, enabling stakeholders to seek recompense for data contributions and fostering a more sustainable collaboration between data providers and model developers. This approach has implications for data provenance, artist rights, and the economics of future AI systems.

Abstract

It is evident that, currently, generative models are surpassed in quality by human professionals. However, with the advancements in Artificial Intelligence, this gap will narrow, leading to scenarios where individuals who have dedicated years of their lives to mastering a skill become obsolete due to their high costs, which are inherently linked to the time they require to complete a task -- a task that AI could accomplish in minutes or seconds. To avoid future social upheavals, we must, even now, contemplate how to fairly assess the contributions of such individuals in training generative models and how to compensate them for the reduction or complete loss of their incomes. In this work, we propose a method to structure collaboration between model developers and data providers. To achieve this, we employ Shapley Values to quantify the contribution of artist(s) in an image generated by the Stable Diffusion-v1.5 model and to equitably allocate the reward among them.
Paper Structure (12 sections, 1 equation, 40 figures, 2 tables)

This paper contains 12 sections, 1 equation, 40 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Image Generation Models
  4. Training Data
  5. Influence evaluation methods
  6. Methodology
  7. Experiments and Results
  8. Evaluation of a single style
  9. Evaluation of multiple styles
  10. Dreambooth
  11. Conclusion
  12. List of prompts

Figures (40)

  • Figure 1: An example of disereable reward split for an image generated by the prompt: an image of space shuttle, by Van Gogh
  • Figure 2: An example of disereable reward split for an image generated by the prompt: an image of a man, by Van Gogh
  • Figure 3: An image generated by the prompt: an image of a Lloyd Shapley
  • Figure 4: Participants (a and c) and the outcome (b) of their interaction for computing Shapley Values.
  • Figure 5: (a) - similarity score between generation, original artworks and pure SD model, (b) - Shapley Values derived from embeddings generated by the CLIP.
  • ...and 35 more figures