Fast Data Attribution for Text-to-Image Models

Sheng-Yu Wang¹

Aaron Hertzmann²

Alexei A. Efros³

Richard Zhang²

Jun-Yan Zhu¹

¹Carnegie Mellon University

²Adobe Research

³UC Berkeley

[Code]

[Paper (NeurIPS 2025)]

Abstract

Data attribution for text-to-image models aims to identify the training images that most significantly influenced a generated output. Existing attribution methods involve considerable computational resources for each query, making them impractical for real-world applications. We propose a novel approach for scalable and efficient data attribution. Our key idea is to distill a slow, unlearning-based attribution method to a feature embedding space for efficient retrieval of highly influential training images. During deployment, combined with efficient indexing and search methods, our method successfully finds highly influential images without running expensive attribution algorithms. We show extensive results on both medium-scale models trained on MSCOCO and large-scale Stable Diffusion models trained on LAION, demonstrating that our method can achieve better or competitive performance in a few seconds, faster than existing methods by 2,500x - 400,000x. Our work represents a meaningful step towards the large-scale application of data attribution methods on real-world models such as Stable Diffusion.

Results

Attribution performance vs. throughput (MSCOCO Models). Previous methods (AbU, D-TRAK) offer high attribution performance but are computationally expensive for deployment. Fast image similarity using off-the-shelf features (DINO) lacks attribution accuracy. We distill slower attribution methods into a feature space that retains attribution performance while enabling fast deployment.

Which feature space is good for attribution? (MSCOCO Models) We compare different feature spaces, before and after tuning for attribution. We measure mAP to the ground truth ranking, generated by AbU+. While text-only embeddings perform well before tuning, image-only embeddings become stronger after tuning. Combining both performs best and is our final method.

Qualitative results (MSCOCO Models). For each generated image and its text prompt on the left, we show top‐5 training images retrieved by: DINO + CLIP‐Text (top row), Ours (middle row), and the ground‐truth influential examples via AbU+ (bottom row). Compared to the untuned baseline, our distilled feature space yields attributions that match the ground‐truth examples more closely.

Results and discussions on Stable Diffuision Models. Below we first show qualitative results for Stable Diffusion models. For each generated image (left), we compare the DINO+CLIP-Text baseline (top row), our calibrated feature ranker (middle row), and AbU+ ground-truth attributions (bottom row). Both AbU+ and our method tend to retrieve images that reflect textual cues rather than visual similarity.

We further verify this phenomenon by checking which feature space predicts attribution results better. In the following figure, while we see similar trends as with MS-COCO, with strongest performing embedding using both text and image features, we find that image-only embeddings perform much worse than text-only embeddings.

It remains challenging to verify whether the "ground‐truth" we collected via AbU+ is accurate or not. Studies by Akyürek et al. and Li et al. suggest that influence‐based methods may weaken on very large models (e.g., LLMs). However, developing more robust attribution algorithms for large‐scale models—and distilling a rank model from stronger teacher methods using our method—are all promising directions ahead.

Paper

Sheng-Yu Wang, Aaron Hertzmann, Alexei A. Efros, Richard Zhang, Jun-Yan Zhu.
Data Attribution for Text-to-Image Models by Unlearning Synthesized Images.
In NeurIPS, 2025. (Paper)

[Bibtex]

Acknowledgements

We thank Simon Niklaus for the help on the LAION image retrieval. We thank Ruihan Gao, Maxwell Jones, and Gaurav Parmar for helpful discussions and feedback on drafts. Sheng-Yu Wang is supported by the Google PhD Fellowship. The project was partly supported by Adobe Inc., the Packard Fellowship, the IITP grant funded by the Korean Government (MSIT) (No. RS-2024-00457882, National AI Research Lab Project), NSF IIS-2239076, and NSF ISS-2403303.

Citation


                @inproceedings{wang2025fastgda,

                  title={Fast Data Attribution for Text-to-Image Models},

                  author={Wang, Sheng-Yu and Hertzmann, Aaron and Efros, Alexei A and Zhang, Richard and Zhu, Jun-Yan},

                  booktitle={NeurIPS},

                  year = {2025},

                  }