Erasure or Erosion? Evaluating Compositional Degradation in Unlearned Text-To-Image Diffusion Models
Arian Komaei Koma, Seyed Amir Kasaei, Ali Aghayari, AmirMahdi Sadeghzadeh, Mohammad Hossein Rohban · 2026
Read Erasure or Erosion? Evaluating Compositional Degradation in Unlearned Text-To-Image Diffusion Models by Arian Komaei Koma, Seyed Amir Kasaei, Ali Aghayari, AmirMahdi Sadeghzadeh, Mohammad Hossein Rohban (2026) 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-2604-04575.txt.
Description
Komaei Koma and colleagues probe unlearning in text-to-image diffusion models via compositional generation, focusing on nudity removal in Stable Diffusion. The team tests varied top unlearning methods with T2I-CompBench++ and GenEval plus set unlearning benchmarks. Results show a steady trade-off between unlearning power and compositional soundness. Strong erasure methods often harm attribute binding, spatial thought and counting, while shape-saving methods often fail at firm erasure. The findings show limits of current test habits and call for unlearning goals that save meaning beyond aimed suppression.
Catalog record
| CREATOR | Arian Komaei Koma, Seyed Amir Kasaei, Ali Aghayari, AmirMahdi Sadeghzadeh, Mohammad Hossein Rohban |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 21369 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.04575