Initialization is Half the Battle: Generating Diverse Images from a Guidance Potential Posterior
Researchers have developed a new method called Diversity-inducing Initialization (DivIn) to address mode collapse in generative AI models. DivIn works by selecting initial noise from a guidance potential posterior, effectively guiding the generation process towards more diverse outputs. This approach is compatible with both diffusion and flow matching models and can be combined with existing trajectory-based methods for even greater improvements in image diversity and quality. AI
IMPACT Enhances diversity in generative models, potentially leading to more varied and creative AI-generated content.