Geometry-Aware Dataset Condensation for Diffusion Model Training

Dataset condensation aims to construct compact datasets from real data via synthesis or selection. However, existing approaches are ill-suited for diffusion model training: synthetic data generation often yields low-fidelity samples unsuitable for authentic modeling, while real subset selection typically fails to preserve the distributional geometry required by diffusion likelihood objectives. To address this, we propose to reformulate real subset selection as a geometry-aware distribution alignment problem. By incorporating one-sided partial optimal transport, our method selectively aligns a compact subset with the full data distribution while allowing unmatched mass in low-density regions, ensuring the preserved geometric structure necessary for effective diffusion model training. To further ensure distributional fidelity, we complement geometric alignment with lightweight feature-statistics and semantic consistency regularization. An efficient two-stage discrete optimization strategy is proposed to achieve this alignment objective. Extensive experiments across diffusion variants, subset sizes, image resolutions, and training rounds show that our method achieves superior fidelity and distributional coverage in diffusion model training.

Modern image-generating AI systems often need very large image collections to learn how to create realistic pictures. This makes training expensive and difficult for researchers or organizations with limited computing resources. A natural solution is to train these systems with a much smaller dataset, but the challenge is choosing which images to keep so that the smaller dataset still represents the original one well. In this paper, we propose a method for selecting a small but useful set of real images for training diffusion models, a popular type of image-generating AI. Instead of creating artificial training images or selecting images with simple ranking rules, our method looks at how the selected images collectively represent the overall structure of the full dataset. It keeps images that cover the important and reliable parts of the data while avoiding less helpful examples. Our experiments show that diffusion models trained on our selected subsets generate better and more diverse images than those trained on subsets chosen by previous methods. This can help reduce the data, storage, and computation needed to train high-quality image generation models.