Saved in:
Bibliographic Details
Main Authors: Gaintseva, Tatiana, Oncescu, Andreea-Maria, Ma, Chengcheng, Liu, Ziquan, Benning, Martin, Slabaugh, Gregory, Deng, Jiankang, Elezi, Ismail
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2503.09630
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910005135933440
author Gaintseva, Tatiana
Oncescu, Andreea-Maria
Ma, Chengcheng
Liu, Ziquan
Benning, Martin
Slabaugh, Gregory
Deng, Jiankang
Elezi, Ismail
author_facet Gaintseva, Tatiana
Oncescu, Andreea-Maria
Ma, Chengcheng
Liu, Ziquan
Benning, Martin
Slabaugh, Gregory
Deng, Jiankang
Elezi, Ismail
contents Diffusion models have transformed image generation, yet controlling their outputs to reliably erase undesired concepts remains challenging. Existing approaches usually require task-specific training and struggle to generalize across both concrete (e.g., objects) and abstract (e.g., styles) concepts. We propose CASteer (Cross-Attention Steering), a training-free framework for concept erasure in diffusion models using steering vectors to influence hidden representations dynamically. CASteer precomputes concept-specific steering vectors by averaging neural activations from images generated for each target concept. During inference, it dynamically applies these vectors to suppress undesired concepts only when they appear, ensuring that unrelated regions remain unaffected. This selective activation enables precise, context-aware erasure without degrading overall image quality. This approach achieves effective removal of harmful or unwanted content across a wide range of visual concepts, all without model retraining. CASteer outperforms state-of-the-art concept erasure techniques while preserving unrelated content and minimizing unintended effects.
format Preprint
id arxiv_https___arxiv_org_abs_2503_09630
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle CASteer: Cross-Attention Steering for Controllable Concept Erasure
Gaintseva, Tatiana
Oncescu, Andreea-Maria
Ma, Chengcheng
Liu, Ziquan
Benning, Martin
Slabaugh, Gregory
Deng, Jiankang
Elezi, Ismail
Graphics
Diffusion models have transformed image generation, yet controlling their outputs to reliably erase undesired concepts remains challenging. Existing approaches usually require task-specific training and struggle to generalize across both concrete (e.g., objects) and abstract (e.g., styles) concepts. We propose CASteer (Cross-Attention Steering), a training-free framework for concept erasure in diffusion models using steering vectors to influence hidden representations dynamically. CASteer precomputes concept-specific steering vectors by averaging neural activations from images generated for each target concept. During inference, it dynamically applies these vectors to suppress undesired concepts only when they appear, ensuring that unrelated regions remain unaffected. This selective activation enables precise, context-aware erasure without degrading overall image quality. This approach achieves effective removal of harmful or unwanted content across a wide range of visual concepts, all without model retraining. CASteer outperforms state-of-the-art concept erasure techniques while preserving unrelated content and minimizing unintended effects.
title CASteer: Cross-Attention Steering for Controllable Concept Erasure
topic Graphics
url https://arxiv.org/abs/2503.09630