Ningning Ding

AutoAnchor: Stable Diffusion Unlearning Using Cross-Attention as a Manifold Surrogate

Siyuan Wen, Jiahao Zeng, Ningning Ding

Diffusion unlearning is essential for mitigating the generation of harmful or copyrighted content in text-to-image models. Current diffusion unlearning techniques determine the model update direction by either using alternatives of the target concept as an anchor or using empty p…Read more