Reproduce “MultiBinder: Preventing Multi-Subject Identity Leakage in Personalized Video 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

Wei Li, Yiheng Zhang, Zhaofan Qiu, Ting Yao, Xiaoyan Sun, Tao Mei

MultiBinder links named subjects to reference identities before arranging their roles in prompts. It uses an entity-aware Q-Former and Entity-Refinement Attention, then a multimodal language model to form role-aware conditioning. On 100-entry MultiBind-Bench, it scores 73.98% IPS and 84.56% VRS, above Kling-1.6’s 69.51% and 82.89%; these automated scores cover entries with 2–4 subjects.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.