Diffusion models have raised concerns about generating illicit content and copyright violations. To address this, researchers have been developing machine unlearning techniques to erase harmful concepts or copyrighted styles from trained models.
The new arXiv preprint, "Compositional Concepts Erasing: Disentangling Object-Style via Domain Manifold Unlearning," proposes a method that specifically targets compositional concepts. The key idea is to disentangle object and style components so that they can be removed independently, rather than erasing the entire concept.
The approach is based on "domain manifold unlearning," which operates on the model's learned representation space. While the abstract does not provide experimental details, the framing suggests a more fine-grained control over what gets erased, potentially reducing collateral damage to unrelated content.