الملخص
Training data attribution for diffusion models aims to identify the training samples that influence a generated instance, but existing methods either require costly per-sample gradient computation or query-specific model optimization. Moreover, most methods attribute changes in a proxy loss rather than changes in the actual model's generative behavior. We address these limitations by formulating attribution directly with a local score discrepancy measure, which applies to any diffusion variant (including DDPM, EDM, and flow matching), and by showing that such measure can be estimated without retraining, as a preconditioned gradient similarity. We instantiate this estimator as Training-data Influence via score Discrepancy (TID), which uses Kronecker-factored curvature to avoid random projections and per-sample gradient storage. We then distill TID into TIDE, a forward-only student trained online to reproduce the teacher's rankings from the diffusion model's internal activations. Under counterfactual evaluation on CIFAR-10, ArtBench-10, and MS-COCO, TID matches or outperforms state-of-the-art approaches, while TIDE retains most of TID's accuracy at four to five orders of magnitude lower per-query cost, attributing generated samples in milliseconds and faster than the generation itself.
الكلمات المفتاحية
الموضوع
بيانات النشر
- المجلة
- غير متاح
- وصول مفتوح
- وصول مفتوح أخضر
اقتبس هذه المقالة
APA 7
Liu, S., Serrà, J., Cheuk, K. W., Kim, J., Choi, W., Ikemiya, Y., Liao, W. H., Ma, J. W., & Mitsufuji, Y. (2026). Distilling Diffusion Score Discrepancy for Efficient Training Data Attribution. https://omanscience.com/ar/articles/distilling-diffusion-score-discrepancy-for-efficient-training-data-attribution
MLA 9
Liu, Shixuan, et al. "Distilling Diffusion Score Discrepancy for Efficient Training Data Attribution." https://omanscience.com/ar/articles/distilling-diffusion-score-discrepancy-for-efficient-training-data-attribution.
شيكاغو (المؤلف–التاريخ)
Liu, Shixuan, Joan Serrà, Kin Wai Cheuk, Jinju Kim, Woosung Choi, Yukara Ikemiya, Wei-Hsiang Liao, Jiaqi W. Ma, and Yuki Mitsufuji. 2026. "Distilling Diffusion Score Discrepancy for Efficient Training Data Attribution." https://omanscience.com/ar/articles/distilling-diffusion-score-discrepancy-for-efficient-training-data-attribution.
هارفارد
Liu, S., Serrà, J., Cheuk, K. W., Kim, J., Choi, W., Ikemiya, Y., Liao, W. H., Ma, J. W. and Mitsufuji, Y. (2026) 'Distilling Diffusion Score Discrepancy for Efficient Training Data Attribution', Available at: https://omanscience.com/ar/articles/distilling-diffusion-score-discrepancy-for-efficient-training-data-attribution.
فانكوفر
Liu S, Serrà J, Cheuk KW, Kim J, Choi W, Ikemiya Y, et al. Distilling Diffusion Score Discrepancy for Efficient Training Data Attribution. https://omanscience.com/ar/articles/distilling-diffusion-score-discrepancy-for-efficient-training-data-attribution
IEEE
S. Liu, J. Serrà, K. W. Cheuk, J. Kim, W. Choi, Y. Ikemiya, W. H. Liao, J. W. Ma, and Y. Mitsufuji, "Distilling Diffusion Score Discrepancy for Efficient Training Data Attribution," https://omanscience.com/ar/articles/distilling-diffusion-score-discrepancy-for-efficient-training-data-attribution.