Post-Generation Curation of Synthetic Images via Homogeneous-Heterogeneous Splitting
The episode discusses the paper "Post-Generation Curation of Synthetic Images via Homogeneous-Heterogeneous Splitting." Hosts explain how using real data structure to guide synthetic image selection, specifically splitting real classes into a Homogeneous and Heterogeneous set, improves model performance. This curation method allows models to learn less obvious variations while significantly reducing the required synthetic data volume for training.