DiverGen Proves AI Models Learn Better with Variety

This section describes DiverGen's comprehensive implementation and visualization techniques. To verify generative diversity, the authors use UMAP visualization and CLIP-based data distribution analysis. While ChatGPT-generated prompts increase textual variety and visual richness, they also improve generative model diversity through the use of Stable Diffusion and DeepFloyd-IF. Compared to previous methods like max CLIP or SAM-foreground, the suggested SAM-background (SAM-bg) annotation method generates more precise and comprehensive masks.