Dataset distillation aims to condense large‐scale datasets into smaller synthetic ones, enabling efficient training with lower storage and computational costs while achieving performance comparable to models trained on the full dataset. Despite notable successes, existing methods often struggle on fine‐grained tasks due to the subtle and complex inter‐class differences inherent in such datasets. To address this challenge, we propose a ControlNet‐enhanced distillation framework that synthesizes high‐quality images with rich fine‐grained details. Specifically, we utilize synthetic images produced by existing distillation methods as supervision signals within a conditional diffusion model guided by ControlNet. To ensure diversity and representativeness, the dataset is partitioned and clustered to construct representative subsets for training and synthesis. ControlNet is fine‐tuned on the training subset using prior knowledge from external synthetic images, and subsequently used to generate high‐resolution images under the guidance of control inputs from the synthesis subset. Experimental results on multiple fine‐grained datasets demonstrate that our method generates diverse, detail‐rich images and achieves state‐of‐the‐art performance. © 2026 Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.
Li et al. (Tue,) studied this question.