Reproduce “Closed-Form Concept Erasure via Double Projections” 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
Chi Zhang, Jingpu Cheng, Zhixian Wang, Ping Liu
The research introduces Concept Erasure with Double Projections (DP), a training-free, closed-form framework that uses two analytical projection steps to remove undesired concepts from generative models. DP effectively suppresses target concepts while maintaining non-target representations, showing a mean preservation drop of 2.4% for object erasure in Stable Diffusion 1.4 and generalizing to FLUX flow-matching models.
Key insights
- Explicitly decoupling erasure and preservation into distinct analytical projection steps effectively minimizes interference with non-target representations.
- The use of a nullspace projection for parameter updates provides theoretical guarantees for the exact preservation of non-target concepts.
- Applying updates directly to embedding layers can mitigate interference from positional embeddings and self-attention, leading to cleaner erasure.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.