Concept Erasure via Attention Redirection

Large text-to-image diffusion models have enabled remarkable creative potential, but they also reproduce copyrighted or sensitive concepts, raising significant ethical and legal concerns. We introduce a lightweight and principled approach for concept erasure, based on a novel mechanism we call attention redirection. Rather than directly suppressing target tokens, we train a dedicated eraser token that actively draws the attention away from the target concept in the image-text attention layers and redirects it toward a neutral anchor token. This intervention prevents the target concept from materializing during image generation, while keeping unrelated content and overall image quality intact. Our training scheme demonstrates strong suppression of the target concept with little collateral damage. Across diverse categories, we outperform existing weight-modification erasure methods while maintaining competitive preservation of non-target content. Moreover, we demonstrate that erasure remains effective even when the model has access to additional context, such as inpainting, where erased concepts often reappear in prior methods. These results establish attention redirection as a more robust and durable mechanism for safe and controllable generative modeling.