SurrogateSHAP Offers Scalable Shapley Attribution for Text-to-Image Diffusion Models

Global: SurrogateSHAP Offers Scalable Shapley Attribution for Text-to-Image Diffusion Models

A team of machine learning researchers has introduced SurrogateSHAP, a framework designed to assign fair value to data contributors in text‑to‑image diffusion models without the need for costly model retraining. The work, posted to arXiv in January 2026, aims to address computational bottlenecks that have limited the practical use of Shapley‑based attribution in generative AI pipelines.

Background on Shapley Attribution in Generative AI

The Shapley value, a concept from cooperative game theory, provides a theoretically sound method for quantifying each participant’s marginal contribution. In the context of diffusion models, applying this metric requires evaluating a combinatorial number of data subsets, which traditionally entails exhaustive retraining—a process that quickly becomes infeasible as dataset sizes grow.

SurrogateSHAP Methodology

SurrogateSHAP circumvents retraining by leveraging inference from a pretrained diffusion model to approximate the utility of each data subset. The approach further employs a gradient‑boosted decision tree to model the utility function, enabling analytical extraction of Shapley values directly from the tree structure. This two‑stage surrogate strategy reduces both computational time and resource consumption.

Experimental Evaluation

The authors assessed the framework across three attribution tasks: image quality for a DDPM‑CFG model on the CIFAR‑20 benchmark, aesthetic scoring for Stable Diffusion on Post‑Impressionist artworks, and product‑diversity measurement for FLUX.1 on a fashion‑product dataset. In each scenario, SurrogateSHAP identified high‑impact contributors more consistently than baseline methods while requiring substantially fewer compute cycles.

Performance Compared to Existing Approaches

Compared with prior attribution techniques that rely on Monte‑Carlo sampling or full‑model retraining, SurrogateSHAP achieved comparable or superior accuracy with up to a 90 % reduction in runtime. The analytical Shapley extraction from the tree model also eliminated variance associated with stochastic sampling.

Case Study: Auditing Clinical Image Generators

In a supplementary analysis, the framework was applied to a clinical image generation pipeline to locate data sources responsible for spurious correlations. The results demonstrated that SurrogateSHAP could pinpoint specific contributors whose inclusion introduced bias, offering a scalable tool for safety audits in high‑stakes AI applications.

Implications for Data Markets and Future Research

By delivering efficient, retraining‑free attribution, SurrogateSHAP paves the way for more transparent data marketplaces where contributors can be compensated based on measurable impact. The authors suggest extending the surrogate model to other generative modalities and exploring integration with existing data licensing frameworks.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.