Conditional Vendi Score: Prompt-Aware Diversity Evaluation for Generative AI Models and LLMs
This research introduces Conditional-Vendi and Conditional RKE scores to measure diversity in text-guided generative models. These new metrics successfully separate model-induced diversity from variability caused by the input prompt. The framework provides a way to quantify how much creativity comes from the model itself versus how much is dictated by the text instructions, offering a tool to guide models toward more diverse outputs.