Entangled Diversity: Rethinking Prompt-Model Separability in Generative Evaluation

Jalali et al. (2024) decompose diversity into prompt-induced and model-induced components, but this assumes separability. We argue that prompts actively shape model diversity—for example, "a diverse crowd" inherently demands high model variance. Using information theory, we’ll introduce an "Entanglement Index" quantifying how prompts constrain or amplify model creativity. Experiments will show that high entanglement correlates with alignment failures in VBench’s spatial-relationship tests. This challenges the core assumption of Conditional Vendi scores and offers a unified lens for diversity-alignment trade-offs.

References:

1. Conditional Vendi Score: An Information-Theoretic Approach to Diversity Evaluation of Prompt-based Generative Models. Mohammad Jalali, Azim Ospanov, Amin Gohari, Farzan Farnia (2024). arXiv.org.
2. VBench: Comprehensive Benchmark Suite for Video Generative Models. Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, Ziwei Liu (2023). Computer Vision and Pattern Recognition.