New Framework Enables Erasure of Thousands of Concepts in Text-to-Image Diffusion Models

The Erasing Thousands of Concepts (ETC) framework provides a scalable solution to safety concerns associated with large-scale text-to-image diffusion models by allowing the elimination of thousands of unwanted concepts while preserving the quality of generated images. Although these models are praised for their visual accuracy, they can inadvertently create problematic content, including copyrighted material. Current methods for concept removal have been constrained to a few hundred concepts due to issues with scalability, precision, and robustness. ETC utilizes a Student's t-distribution Mixture Model (tMM) to model low-rank concept distributions, enabling targeted concept removal through affine optimal transport. Furthermore, it incorporates a Mixture-of-Experts (MoE)-based module named MoEr to bolster its functionality. This research, available in arXiv preprint 2604.16481v1, marks a notable step forward in enhancing the practicality and effectiveness of concept erasure.