Reproduce “ID-Patch: Robust ID Association for Group Photo Personalization” 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
Yimeng Zhang, Tiancheng Zhi, Jing Liu, Shen Sang, Liming Jiang, Qing Yan, Sijia Liu, Linjie Luo
ID-Patch introduces a dual mechanism using identity-specific patches for precise spatial control and ID embeddings for detailed facial resemblance, enabling robust personalization of multiple identities in group photos. This method achieves higher identity resemblance (0.751 FaceNet similarity) and positioning accuracy (0.958) while drastically reducing generation time to 9.69 seconds, which remains near-constant with increasing numbers of people.
Key insights
- Encoding identity information through a dual mechanism (spatial patches for position and detailed embeddings for resemblance) from shared facial features effectively disentangles and robustly associates identity with location.
- A two-stage training approach is crucial for establishing distinct ID patches and robust spatial control before refining detailed facial resemblance via ID embeddings.
- ID patches can seamlessly integrate with existing spatial conditioning models like ControlNet by direct overlay, enhancing control without adding architectural complexity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.