Reproduce “ObjectMate: A Recurrence Prior for Object Insertion and Subject-Driven 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
Daniel Winter, Asaf Shul, Matan Cohen, Dana Berman, Yael Pritch, Alex Rav-Acha, Yedid Hoshen
ObjectMate introduces a method to create high-quality, tuning-free object compositions by automatically building a large supervised dataset from naturally recurring object instances in unlabeled web images. The system demonstrates superior photorealistic integration and object identity preservation in both object insertion and subject-driven generation tasks, achieving user preferences for realism (80-86%) and identity (76-100%) in insertion.
Key insights
- Many everyday, mass-produced objects naturally recur in large internet datasets, providing a scalable and diverse source of multi-view training data.
- Instance retrieval (IR) features are crucial for identifying exact object instances for identity preservation, outperforming general semantic features.
- A diffusion model trained on naturally occurring recurrent object data can achieve high-fidelity, tuning-free object composition, addressing key limitations of prior approaches.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.