Choose Your Anchor Wisely: Effective Unlearning Diffusion Models via Concept Reconditioning

Large-scale conditional diffusion models (DMs) have demonstrated exceptional ability in generating high-quality images from textual descriptions, gaining widespread use across various domains. However, these models also carry the risk of producing harmful, sensitive, or copyrighted content, creating a pressing need to remove such information from their generation capabilities. While retraining from scratch is prohibitively expensive, machine unlearning provides a more efficient solution by selectively removing undesirable knowledge while preserving utility. In this paper, we introduce \textbf{COncept REconditioning (CORE)}, a simple yet effective approach for unlearning diffusion models. Similar to some existing approaches, CORE guides the noise predictor conditioned on forget concepts towards an anchor generated from alternative concepts. However, CORE introduces key differences in the choice of anchor and retain loss, which contribute to its enhanced performance. We evaluate the unlearning effectiveness and retainability of CORE on UnlearnCanvas. Extensive experiments demonstrate that CORE surpasses state-of-the-art methods including its close variants and achieves near-perfect performance, especially when we aim to forget multiple concepts. More ablation studies show that CORE's careful selection of the anchor and retain loss is critical to its superior performance.