Reproduce “IdentityStory: Taming Your Identity-Preserving Generator for Human-Centric Story Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Donghao Zhou, Jingyu Lin, Guibao Shen, Quande Liu, Jialin Gao, Lihao Liu, Lan Du, Cunjian Chen and 3 more
Researchers from The Chinese University of Hong Kong, Monash University, and collaborators developed IdentityStory, a framework for generating human-centric visual stories with consistent characters from text prompts. The method achieved a 55.5% Face-Sim score, over double that of prior state-of-the-art methods, on a new human-centric benchmark while maintaining strong text alignment and image quality.
Key insights
- Identity-preserving generators have a structured identity space that can be effectively utilized to derive unique character identities from text-only descriptions via iterative filtering.
- Mitigating text alignment degradation in identity-preserving models is achievable by injecting identities during the re-denoising process, using contextual information from a general T2I model and layout guidance.
- Decoupling identity extraction from scene generation, and employing progressive masking for character integration, allows for precise identity preservation within diverse narrative contexts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.