Exploring More to Solve More: Boosting Diversity in Text Diffusion Models via Entropy-Based Guidance

Semantic-Aware Kernel Entropy (SAKE) guidance increases text diffusion diversity by computing order-2 Renyi entropy over a kernel Gram matrix. This training-free method linearizes semantic interactions and token positions in the embedding space to dynamically adjust sampling distributions between exploration and fidelity.