Empty SPACE: Cross-Attention Sparsity for Concept Erasure in Diffusion Models
Nicola Novello, Andrea M. Tonello · 2026
Read Empty SPACE: Cross-Attention Sparsity for Concept Erasure in Diffusion Models by Nicola Novello, Andrea M. Tonello (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-2605-10198.txt.
Description
Novello and Tonello propose SPACE, a method for erasing concepts from text-to-image diffusion models without retraining. Erasure matters for avoiding copyrighted and explicit content, and closed-form methods offer a quick alternative to backpropagation-based techniques. But closed-form erasure weakens when scaling from smaller models like Stable Diffusion 1.5 to larger ones like Stable Diffusion XL. SPACE, short for sparse cross-attention-based concept erasure, restores effectiveness at larger scale.
The method iteratively modifies cross-attention parameters with a closed-form update that jointly creates sparsity and erases target concepts. Concentrating the concept mapping into a lower-dimensional subspace beats dense baselines on erasure efficacy and robustness against adversarial prompts. SPACE also reaches eighty to ninety percent cross-attention sparsity, cutting storage for modified parameters by seventy percent, a memory efficiency win alongside the safety gain. The paper warns that it contains model outputs that may offend.
Catalog record
| CREATOR | Nicola Novello, Andrea M. Tonello |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 58 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.10198