Erasing Thousands of Concepts: Towards Scalable and Practical Concept Erasure for Text-to-Image Diffusion Models

A team of researchers has introduced a breakthrough framework called 'Erasing Thousands of Concepts' (ETC) that tackles a critical safety and copyright problem in text-to-image AI. Current methods for removing unwanted concepts—like copyrighted characters or unsafe imagery—from models like Stable Diffusion are limited, struggling to erase more than a few hundred items without degrading overall image quality. The ETC framework scales this capability by an order of magnitude, successfully erasing over 2,000 concepts while maintaining generation fidelity.

The technical innovation lies in a two-stage process. First, it models concept distributions using a Student's t-distribution Mixture Model (tMM), which allows for precise 'pin-point' removal of target concepts via affine optimal transport while anchoring and preserving unrelated ones. Second, it trains a specialized Mixture-of-Experts module dubbed 'MoEraser' to surgically remove the target concept embeddings. Crucially, the team also engineered the system for robustness by injecting noise into the model's text embedding projector and fine-tuning MoEraser for recovery, making it resistant to white-box attacks that try to bypass the erasure.

Extensive testing across heterogeneous domains and multiple diffusion models demonstrates state-of-the-art performance. This research, published on arXiv, represents a significant leap towards making powerful generative AI models safer and more legally compliant for widespread use, moving from proof-of-concept removals to systematic, large-scale content filtering.