Dynamic Eraser for Guided Concept Erasure in Diffusion Models

TL;DR

This paper introduces Dynamic Semantic Steering (DSS), a training-free framework for precise concept erasure in diffusion models, significantly improving safety and content control in text-to-image generation.

Contribution

DSS automates semantic anchor discovery and leverages cross-attention features for controllable, interpretable concept erasure without training, outperforming existing methods.

Findings

Achieves an average erasure rate of 91.0%

Outperforms SOTA methods with erasure rates from 18.6% to 85.9%

Minimal impact on output fidelity

Abstract

Concept erasure in Text-To-Image (T2I) diffusion models is vital for safe content generation, but existing inference-time methods face significant limitations. Feature-correction approaches often cause uncontrolled over-correction, while token-level interventions struggle with semantic granularity and context. Moreover, both types of methods are prone to severe semantic drift or even complete representation collapse. To address these challenges, we present Dynamic Semantic Steering (DSS), a lightweight, training-free framework for interpretable and controllable concept erasure. DSS introduces: 1) Sensitive Semantic Boundary Modeling (SSBM) to automate the discovery of safe semantic anchors, and 2) Sensitive Semantic Guidance (SSG), which leverages cross-attention features for precise detection and performs correction via a closed-form solution derived from a well-posed objective. This ensures optimal suppression of sensitive content while preserving benign semantics. DSS achieves an average erasure rate of 91.0\%, significantly outperforming SOTA methods (from 18.6\% to 85.9\%) with minimal impact on output fidelity.