On July 2, 2026, researchers Disheng Liu, Tuo Liang, Chaoda Song, and Yu Yin introduced a novel approach to enhance the utility of synthetic images in machine learning. Their study, titled Post-Generation Curation of Synthetic Images via Homogeneous-Heterogeneous Splitting, reveals how effective selection of generated images can significantly improve downstream performance without retraining the generator.
Addressing the Limitations of Generative Models
Recent advancements in generative models have enabled the production of high-quality synthetic images, which serve as valuable training data for data-intensive models. However, existing methods often rely on either fine-tuning generators or implementing lightweight adaptations, making them generator-specific and requiring expert knowledge. Liu and colleagues explored a different angle by asking whether simply selecting a more informative subset of generated images could enhance their utility.
The researchers found that modern generators exhibit a structural bias, producing an excess of canonical modes while neglecting intra-class variation. By recognizing this bias, they proposed a method that categorizes real classes into two subsets: a canonical Homogeneous (HO) subset and a non-redundant Heterogeneous (HE) subset. This innovative selection process aims to counteract the limitations of conventional generative models.
Methodology and Results
The proposed method scores synthetic images based on a fidelity-diversity criterion, which rewards semantic alignment while penalizing redundancy among canonical samples. This generator-agnostic approach does not require any retraining, making it highly efficient. In experiments across multiple benchmarks, the method consistently outperformed state-of-the-art data selection baselines.
