<ul data-eligibleForWebStory="true"><li>Diffusion models have been successful in high-fidelity image synthesis and prompt-guided generative modeling.</li><li>Ensuring diversity in generated samples of prompt-guided diffusion models is a challenge, especially with broad semantic prompts.</li><li>Recent methods introduced diversity measures for varied generations.</li><li>The SPARKE method is proposed for prompt-aware diversity guidance, using conditional entropy.</li><li>SPARKE enables prompt-aware diversity control by dynamically conditioning diversity measurement on similar prompts.</li><li>Entropy computation in SPARKE poses challenges in large-scale generation, but the method focuses on Conditional latent RKE Score Guidance to address this.</li><li>The reduced computation complexity in SPARKE allows for diversity-guided sampling over many generation rounds on different prompts.</li><li>Numerical testing on text-to-image diffusion models shows that SPARKE improves prompt-aware diversity without significant computational costs.</li><li>The code for the SPARKE method is available on the project page: https://mjalali.github.io/SPARKE</li></ul>

SPARKE: Scalable Prompt-Aware Diversity Guidance in Diffusion Models via RKE Score

Discover more