New method curates synthetic images, matches real data with 40% fewer samples

A team led by Disheng Liu has introduced a novel post-generation curation technique called Homogeneous-Heterogeneous (HO-HE) Splitting. The approach addresses a fundamental bias in modern generative models: they tend to over-produce canonical modes of each class while underrepresenting intra-class variation. By splitting real training classes into a Homogeneous (HO) subset—images near the canonical mode—and a Heterogeneous (HE) subset—non-redundant, diverse examples—the method constructs a reference distribution. Synthetic images are then scored using a fidelity-diversity criterion that rewards semantic alignment with real data while penalizing redundancy relative to the HO subset.

This curation step is completely generator-agnostic and requires no fine-tuning or prompt engineering. In experiments across multiple benchmarks, HO-HE Splitting consistently outperformed state-of-the-art data selection baselines. Remarkably, it enabled downstream models to match real-data performance with up to 40% fewer synthetic samples. The technique also stack on top of stronger task-tuned generators, yielding additional gains on both classification and segmentation tasks. The paper demonstrates that intelligent selection after generation is a powerful, complementary mechanism to improve synthetic data utility—not a replacement for better generators.