Towards Consistent and Efficient Dataset Distillation via Diffusion-Driven Selection

Researchers have proposed a novel framework for dataset distillation using diffusion-driven selection, achieving consistent gains over state-of-the-art methods in image classification benchmarks.

The proposed framework utilizes pre-trained diffusion models for patch selection rather than generation, bypassing pixel-level optimization^[1]. By predicting noise from the diffusion model conditioned on input images and optional text prompts, the method identifies distinctive regions within original images. Intra-class clustering and ranking are applied to enforce diversity constraints on selected patches^[1]. According to arXiv^[1] submission records, the initial version of this research was submitted on 13 Dec 2024, with subsequent revisions culminating in the final version on 25 Jun 2026. A related study on arxiv.org^[2] noted that most existing diffusion-based methods adopt a rigid 'Generate-and-Use' strategy, whereas the proposed framework decouples generation, selection, and refinement, enabling more effective use of the distillation budget. Experiments on large-scale and fine-grained image classification benchmarks showed consistent gains over diffusion-based baselines^[2].