SurrogateSHAP: Training-Free Contributor Attribution for Text-to-Image (T2I) Models

As Text-to-Image (T2I) diffusion models are increasingly used in real-world creative workflows, a principled framework for valuing contributors who provide a collection of data is essential for fair compensation and sustainable data marketplaces. While the Shapley value offers a theoretically grounded approach to attribution, it faces a dual computational bottleneck: (i) the prohibitive cost of exhaustive model retraining for each sampled subset of players (i.e., data contributors) and (ii) the combinatorial number of subsets needed to estimate marginal contributions due to contributor interactions. To this end, we propose SurrogateSHAP, a retraining-free framework that approximates the expensive retraining game through inference from a pretrained model. To further improve efficiency, we employ a gradient-boosted tree to approximate the utility function and derive Shapley values analytically from the tree-based model. We evaluate SurrogateSHAP across three diverse attribution tasks: (i) image quality for DDPM-CFG on CIFAR-20, (ii) aesthetics for Stable Diffusion on Post-Impressionist artworks, and (iii) product diversity for FLUX.1 on Fashion-Product data. Across settings, SurrogateSHAP outperforms prior methods while substantially reducing computational overhead, consistently identifying influential contributors across multiple utility metrics. Finally, we demonstrate that SurrogateSHAP effectively localizes data sources responsible for spurious correlations in clinical images, providing a scalable path toward auditing safety-critical generative models. Code is available at https://anonymous.4open.science/r/CFG-Attribution-15DD/

As text-to-image (T2I) generative models are increasingly used in real-world applications, it is crucial to determine how much value individual creators contribute to the training data so they can be fairly compensated. The gold standard for solving this credit assignment problem is the Shapley value, which works by removing contributors one at a time to measure how the model's output changes. However, applying this to massive T2I models is currently impractical. To measure exactly how much value a specific piece of data adds, the model must be completely retrained for every possible combination of data, making the process incredibly slow and expensive. To solve this, we introduce SurrogateSHAP. First, instead of training or retraining the model every time, we show that the original, fully-trained model already contains valuable information about these individual data contributions. By leveraging this insight, our framework takes a clever shortcut using test-time computation. Second, we approximate this exhaustive data-removal process using a fast, tree-based algorithm to efficiently estimate values without exhaustively trying all combinations. Our findings show that this approach is 2.9 to 23.4 times more efficient than existing methods. In our tests across diverse tasks—from evaluating image quality and artistic aesthetics to measuring fashion product diversity—SurrogateSHAP identified important data contributors much more effectively. Beyond ensuring creators receive fair credit, our method also serves as an auditing tool for clinical data. When applied to a clinical dermatology T2I model, it tracked down the exact hospital data sources responsible for biased outputs in the generated skin images, providing a clear path toward making these AI models safer and more transparent.