Reproduce “Erasing Without Collateral Damage: Precise Concept Removal in Diffusion Models” 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
Parth Upman, Nishita Jain, Shreyank N Gowda
CARE (Covariance-Aware Retained-subspace Erasure) is a training-free method that precisely removes target concepts from text-to-image diffusion models by employing a covariance-aware erasure direction in the cross-attention value space. This approach significantly reduces collateral damage to semantically related non-target concepts, achieving improved preservation across various concept types with minimal computational overhead.
Key insights
- Collateral damage in concept erasure arises from existing methods conflating target identity with shared visual structures common to related concepts.
- Explicitly modeling retained concepts through a covariance-aware direction allows for a more precise separation of target-specific information from general generative knowledge.
- A continuous shrinkage parameter (γ) offers practical control over the erase-preserve trade-off, enabling customization of erasure strength versus preservation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.