Label-Consistent Dataset Distillation with Detector-Guided Refinement

TL;DR

Experimental results demonstrate that the proposed detector-guided dataset distillation framework can synthesize high-quality representative images with richer details, achieving state-of-the-art performance on the validation set.

Abstract

Dataset distillation (DD) aims to generate a compact yet informative dataset that achieves performance comparable to the original dataset, thereby reducing demands on storage and computational resources. Although diffusion models have made significant progress in dataset distillation, the generated surrogate datasets often contain samples with label inconsistencies or insufficient structural detail, leading to suboptimal downstream performance. To address these issues, we propose a detector-guided dataset distillation framework that explicitly leverages a pre-trained detector to identify and refine anomalous synthetic samples, thereby ensuring label consistency and improving image quality. Specifically, a detector model trained on the original dataset is employed to identify anomalous images exhibiting label mismatches or low classification confidence. For each defective image, multiple candidates are generated using a pre-trained diffusion model conditioned on the corresponding image prototype and label. The optimal candidate is then selected by jointly considering the detector's confidence score and dissimilarity to existing qualified synthetic samples, thereby ensuring both label accuracy and intra-class diversity. Experimental results demonstrate that our method can synthesize high-quality representative images with richer details, achieving state-of-the-art performance on the validation set.

Figures (5)

• Figure 1: Visualization comparison of different dataset distillation methods. CR shows distilled images obtained by the conventional method SRe$^{2}$L yin2023sre2l, while GR denotes images synthesized by the diffusion-based generative method D$^{4}$M su2024d.
• Figure 2: Overview of the proposed framework. It comprises two primary modules: prototype-guided image synthesis and anomaly detection with iterative refinement. The first module synthesizes images guided by class-specific prototypes using a diffusion model. In contrast, the second module identifies defective samples, which are refined through candidate generation and selection by jointly considering the detector’s confidence and the dissimilarity to existing qualified samples.
• Figure 3: Comparison of Grad-CAM visualizations across different methods. The first row shows the original input images. The second row presents attention maps generated by D4M, while the third row shows results from our proposed method.
• Figure 4: Visualization of images generated using different semantic strategies. For each column, the images are generated using the same image prototype and random seed. The left panel displays results on ImageNette ($256 \times 256$ pixels), and the right panel shows results on CIFAR-10 ($32 \times 32$ pixels). Compared to the baseline, our approach generates images with more complete semantic structures and clearer class-discriminative features.
• Figure 5: Parameter Analysis. (a) Effect of the top-$k$ candidate pool size on accuracy. (b) Effect of the softmax confidence threshold $\beta$ on accuracy.