The authors use sparse autoencoders, tools that decompose internal neural-network activations into a small set of active features, to identify features associated with a target concept.
Targeted feature edits improve concept removal from image diffusion models
ConceptACP uses sparse autoencoders to guide localized fine-tuning, reporting a 36.34-percentage-point average improvement in forgetting success on SDXL-Turbo.
Research Lab
Yiran Xiahou · Siliang Qin · Kai Chen · Shengzhi Zhang · Renyang Liu
Institute of Information Engineering, Chinese Academy of Sciences · Boston University · NationalUniversity of SingaporeSingapore
Research Digest··2 min read
Xiahou and colleagues introduce Concept Acupuncturist, a method for suppressing selected concepts in text-to-image diffusion models by fine-tuning one internal block.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.