Dynamic Semantic Steering Framework Enables Precise Concept Erasure in Diffusion Models
A novel framework, Dynamic Semantic Steering (DSS), has been developed to tackle significant challenges in concept erasure for Text-To-Image diffusion models without the need for training. Current methods often lead to excessive corrections or difficulties with semantic precision, resulting in semantic drift or representation collapse. DSS features Sensitive Semantic Boundary Modeling (SSBM) to identify secure semantic anchors automatically. Additionally, it incorporates Sensitive Semantic Guidance (SSG), which leverages cross-attention features for accurate detection and utilizes a closed-form solution from a well-defined objective for corrections. This method guarantees effective suppression of sensitive content while preserving the integrity of the model. Concept erasure is crucial for safe AI content generation. The framework is lightweight and aims for interpretable, controllable functionality. The study was published on arXiv with the identifier 2604.16483v1.
Key facts
- Dynamic Semantic Steering (DSS) is a training-free framework for concept erasure in Text-To-Image diffusion models
- Existing methods face limitations like uncontrolled over-correction and semantic drift
- DSS introduces Sensitive Semantic Boundary Modeling (SSBM) to automate discovery of safe semantic anchors
- Sensitive Semantic Guidance (SSG) leverages cross-attention features for precise detection
- Correction is performed via a closed-form solution from a well-posed objective
- Concept erasure is vital for safe content generation in diffusion models
- The research was published on arXiv under identifier 2604.16483v1
- The framework is lightweight and designed for interpretable and controllable operation
Entities
Institutions
- arXiv