Reproduce “Compositional Concepts Erasing: Disentangling Object-Style via Domain Manifold Unlearning” 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

Yang Xiao, Kaiyuan Deng, Gen Li, Zinan Ling, Zhiwen Fan, Xiaolong Ma, Bo Hui

DMU selectively unlearns a target object in specified styles while retaining it in allowed styles and preserving other objects. It uses retained-style features of the target object as a reference, repelling features from styles to unlearn. On UnlearnCanvas, it reaches 85.81 HA versus CUIG’s 70.49 across six domains; tests cover one target object at a time.

Summary by alphaXiv.

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