Post-Generation Curation of Synthetic Images via Homogeneous-Heterogeneous Splitting

Short summary

Researchers propose a post-generation curation method for synthetic images that splits each class into canonical and diverse subsets, then scores samples by fidelity-diversity criteria. The approach achieves real-data performance with up to 40% fewer synthetic samples across classification and segmentation tasks. This generator-agnostic method complements stronger models without requiring retraining.

• •Splits synthetic image classes into homogeneous (canonical) and heterogeneous (diverse) subsets to counter generator bias
• •Achieves real-data performance with 40% fewer synthetic samples via fidelity-diversity scoring
• •Generator-agnostic post-selection method that works with any generative model without retraining

Generated with AI, which can make mistakes.