Reproduce “Infinite-ID: Identity-preserved Personalization via ID-semantics Decoupling Paradigm” 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
Yi Wu, Ziqiang Li, Heliang Zheng, Chaoyue Wang, Bin Li
Infinite-ID presents a novel ID-semantics Decoupling Paradigm for personalized text-to-image generation that resolves the inherent trade-off between identity fidelity and semantic consistency. The method achieves superior identity preservation and text adherence by training on identity information separately and integrating it with text features during inference via a mixed attention mechanism.
Key insights
- Decoupling identity and semantic information during training and recombining them during inference resolves the trade-off between identity fidelity and semantic consistency in text-to-image personalization.
- Training exclusively on identity preservation without text prompts forces the model to learn robust identity features without semantic interference.
- A mixed attention mechanism and AdaIN-mean operation effectively integrate separately learned identity and semantic features during inference for controlled generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.