IMS3: Breaking Distributional Aggregation in Diffusion-Based Dataset Distillation
Published in arXiv, 2026
This paper addresses a core limitation in diffusion-based dataset distillation: while these methods generate diverse samples, they optimize for generative likelihood rather than classification performance. The authors propose two complementary solutions: Inversion-Matching refines the denoising process to broaden coverage and variety, and Selective Subgroup Sampling improves class separation by identifying representative yet distinctive synthetic samples.
Recommended citation: Chenru Wang, Yunyi Chen, Zijun Yang, Joey Tianyi Zhou, Chi Zhang. (2026). "IMS3: Breaking Distributional Aggregation in Diffusion-Based Dataset Distillation." arXiv preprint arXiv:2603.13960.
Download Paper