Zero-Residual Concept Erasure via Progressive Alignment in Text-to-Image Model

Hongxu Chen, Zhen Wang, Taoran Mei, Lin Li, Bowei Zhu, Runshi Li, Long Chen · 2025

Read Zero-Residual Concept Erasure via Progressive Alignment in Text-to-Image Model by Hongxu Chen, Zhen Wang, Taoran Mei, Lin Li, Bowei Zhu, Runshi Li, Long Chen (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-2508-04472.txt.

Description

Image generators can be taught to forget a harmful concept, but current fixes leave traces behind and dull the model's craft. Hongxu Chen and co-authors name two flaws. First, alignment between the banned concept and a safe stand-in is never exact, so scraps remain. Second, updates pile into a few deep layers, which warps other output.

Their method, ErasePro, forces an exact zero-residual match in the math. It also spreads the change gradually from shallow layers to deep ones, so deep layers barely move. Across erasing objects, art styles, and nudity, the approach removed targets more fully while keeping image quality. Earlier closed-form methods missed that balance.

Catalog record

CREATORHongxu Chen, Zhen Wang, Taoran Mei, Lin Li, Bowei Zhu, Runshi Li, Long Chen
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE40 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-04472

Browse the full library · About the Tower of Babel