Reproduce “CustAny: Customizing Anything from A Single Example” 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
Lingjie Kong, Kai Wu, Xiaobin Hu, Wenhui Han, Jinlong Peng, Chengming Xu, Donghao Luo, Mengtian Li and 3 more
Researchers from Fudan University and Tencent Youtu Lab developed CustAny, a zero-shot framework capable of customizing diverse objects from a single reference image with high identity fidelity and text editability. The method, leveraging a newly constructed large-scale ID-consistent dataset, achieves improved performance over existing general and specialized customization techniques.
Key insights
- Robust identity extraction for diverse objects benefits from combining complementary self-supervised models like DINOv2 and MAE, capturing both intricate details and color consistency.
- Effective identity preservation while maintaining text editability requires a dual-level injection strategy, integrating ID information at both semantic (global) and fine-grained (local) levels within the diffusion model.
- Training with an ID-aware decoupling module helps disentangle identity features from non-ID attributes, leading to greater generation diversity and more precise text control.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.