Text-to-image models face several safety concerns that may hinder their applicability for deployment. Traditionally, these concerns, including bias, copyright infringement, and offensive content, have been addressed separately. However, in real-world scenarios, all these issues can arise simultaneously within the same model. To overcome this challenge, we introduce a novel approach called Unified Concept Editing (UCE), which efficiently handles all these problems using a single method.
Our UCE method does not require retraining the model and instead employs a closed-form solution for editing. Additionally, it seamlessly scales to accommodate concurrent edits on text-conditional diffusion models. Through our approach, we demonstrate the ability to debias, erase stylistic elements, and moderate content simultaneously in text-to-image projections. Extensive experiments validate the improved effectiveness and scalability of our method compared to previous approaches.
For those interested, our code is readily available at https://unified.baulab.info.