Concept Acupuncturist: An Efficient and Precise SAE-Based Concept Erasure Method for Diffusion Models

The researchers propose Concept Acupuncturist (ConceptACP), a diffusion model machine unlearning framework that combines interpretability analysis with SAEs, for efficiently and precisely erasing harmful or violating concepts. The method establishes a mapping between semantic concepts and the cross-attention blocks in the U-Net denoiser, locates the blocks most relevant to the target concept for targeted optimization, and avoids large-scale fine-tuning over the full diffusion trajectory; at the same time, it uses the sparse, semantic feature activations obtained from the SAE to construct a correction signal, which is projected back to the output of the corresponding block as supervision to alleviate concept entanglement, and therefore requires no external inference-time module. Under the standard single-step generation setting of SDXL-Turbo, ConceptACP's unlearning success rate improves by an average of 36.34 percentage points over the best baseline, while maintaining image quality and prompt fidelity, and can effectively generalize to other Stable Diffusion backbones. Code is open-sourced at https://github.com/Eran-Xia/ConceptACP。