Abstract:Text-to-image diffusion models have demonstrated the underlying risk of generating various unwanted content, such as sexual elements. To address this issue, the task of concept erasure has been introduced, aiming to erase any undesired concepts that the models can generate. Previous methods, whether training-based or training-free, have primarily focused on the input side, i.e. texts. However, they often suffer from incomplete erasure due to limitations in the generalization from limited prompts to diverse image content. In this paper, motivated by the notion that concept erasure on the output side, i.e. generated images, may be more direct and effective, we propose to check concepts based on intermediate-generated images and correct them in the remainder of the generation process. Two key challenges are identified, i.e. determining the presence of target concepts during generation and replacing them on the fly. Leveraging the generation mechanism of diffusion models, we present the Concept Corrector, which incorporates the Generation Check Mechanism and the Concept Removal Attention. This method can identify the generated features associated with target concepts and replace them using pre-defined negative prompts, thereby achieving concept erasure. It requires no changes to model parameters and only relies on a given concept name and its replacement content. To the best of our knowledge, this is the first erasure method based on intermediate-generated images. The experiments on various concepts demonstrate its impressive erasure performance. Code: https://github.com/RichardSunnyMeng/ConceptCorrector.
Abstract:Text-to-image diffusion models have been demonstrated with unsafe generation due to unfiltered large-scale training data, such as violent, sexual, and shocking images, necessitating the erasure of unsafe concepts. Most existing methods focus on modifying the generation probabilities conditioned on the texts containing unsafe descriptions. However, they fail to guarantee safe generation for unseen texts in the training phase, especially for the prompts from adversarial attacks. In this paper, we re-analyze the erasure task and point out that existing methods cannot guarantee the minimization of the total probabilities of unsafe generation. To tackle this problem, we propose Dark Miner. It entails a recurring three-stage process that comprises mining, verifying, and circumventing. It greedily mines embeddings with maximum generation probabilities of unsafe concepts and reduces unsafe generation more effectively. In the experiments, we evaluate its performance on two inappropriate concepts, two objects, and two styles. Compared with 6 previous state-of-the-art methods, our method achieves better erasure and defense results in most cases, especially under 4 state-of-the-art attacks, while preserving the model's native generation capability. Our code will be available on GitHub.