IdentityStory: Human-Centric Story Generation with Consistent Characters

This paper addresses the challenge of generating stories with consistent human characters in visual generative models. It introduces IdentityStory, a framework designed to maintain detailed face consistency and coordinate multiple characters across sequential images. The key contributions are Iterative Identity Discovery and Re-denoising Identity Injection, which aim to improve character identity preservation. The paper's significance lies in its potential to enhance the realism and coherence of human-centric story generation, particularly in applications like infinite-length stories and dynamic character composition.