Curating Synthetic Images Improves Downstream AI Model Utility.

A new study proposes a method for improving the utility of synthetic images for training AI models, focusing on post-generation curation rather than generator fine-tuning. Modern generative models often produce high-quality images but tend to overemphasize common variations within a class while underrepresenting diverse, less canonical examples. This structural bias can limit the effectiveness of synthetic data. The researchers introduce Homogeneous-Heterogeneous (HO-HE) splitting, which divides real data into canonical (homogeneous) and non-redundant (heterogeneous) subsets. Synthetic images are then scored based on a fidelity-diversity criterion that rewards semantic alignment with real data while penalizing redundancy in canonical modes. This generator-agnostic approach consistently outperforms existing data selection methods, achieving comparable performance to real data with significantly fewer synthetic samples and proving effective even with stronger, task-tuned generators.