Title: SurrogateSHAP: A Training-Free Approach to Contributor Attribution in Text-to-Image Models

Abstract:

As Text-to-Image (T2I) diffusion models become deeply integrated into practical creative processes, establishing a rigorous framework for valuing data contributors is crucial for ensuring equitable compensation and fostering sustainable data markets. Although the Shapley value provides a robust theoretical basis for attribution, its application is hindered by two significant computational challenges: the excessive cost of retraining models for every sampled subset of contributors, and the combinatorial explosion of subsets required to accurately estimate marginal contributions resulting from interactions among contributors.

To address these issues, we introduce SurrogateSHAP, a novel framework that eliminates the need for retraining by approximating the costly retraining game through inference on a pre-existing model. To further enhance efficiency, we utilize a gradient-boosted tree to model the utility function, allowing for the analytical derivation of Shapley values from this tree-based structure.

We assess SurrogateSHAP’s performance across three distinct attribution scenarios: evaluating image quality using DDPM-CFG on the CIFAR-20 dataset, assessing aesthetics with Stable Diffusion on Post-Impressionist artworks, and measuring product diversity via FLUX.1 on Fashion-Product data. Our results demonstrate that SurrogateSHAP not only surpasses existing methods but also significantly lowers computational demands, reliably pinpointing key contributors across various utility metrics. Additionally, we show that SurrogateSHAP can effectively identify the data sources driving spurious correlations in clinical images, offering a scalable solution for auditing the safety of critical generative models.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC