Fantastic Targets for Concept Erasure in Diffusion Models and Where To Find Them
Anh Bui, Trang Vu, Long Vuong, Trung Le, Paul Montague, Tamas Abraham, Junae Kim, Dinh Phung · 2025
Read Fantastic Targets for Concept Erasure in Diffusion Models and Where To Find Them by Anh Bui, Trang Vu, Long Vuong, Trung Le, Paul Montague, Tamas Abraham, Junae Kim, Dinh Phung (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-18950.txt.
Description
Anh Bui and co-authors rethink how diffusion image models are made to forget a concept. Erasure guards against harmful outputs by unlearning a chosen idea. The standard trick maps the banned idea onto one fixed neutral stand-in, the same for every case.
Their 2025 paper shows that fixed aims can wound nearby ideas. By laying the concept space out as a graph, they find that fallout from an erasure stays local. Their method, Adaptive Guided Erasure, picks the best target for each banned concept on the fly. Tests beat leading erasure methods at keeping unrelated concepts whole while the banned one fades. The results recast erasure as a routing problem, not a blunt overwrite.
Catalog record
| CREATOR | Anh Bui, Trang Vu, Long Vuong, Trung Le, Paul Montague, Tamas Abraham, Junae Kim, Dinh Phung |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 90 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-18950