arXiv · 2611.20625
ASCG: Adaptive Spatial Classifier Guidance for Surgical Concept Suppression in Diffusion Models
Abstract
While text-to-image (T2I) diffusion models achieve remarkable quality, they risk producing harmful content. Existing safety mechanisms often rely on indiscriminate, global interventions that induce unintended semantic drift and degrade image quality. To address this, we propose Adaptive Spatial Classifier Guidance (ASCG), a necessity-driven framework that employs probabilistic monitoring to activate guidance selectively upon harm detection. By leveraging Grad-CAM for localized suppression, ASCG effectively removes harmful concepts while preserving superior semantic fidelity compared to conventional methods. Furthermore, we overcome the limitations of traditional metrics like NudeNet---which fails to assess semantic integrity---by introducing a VLM-based evaluation framework to provide a fine-grained, robust analysis of both safety and context preservation. Our results demonstrate that ASCG not only balances safety and fidelity but also exhibits remarkable plug-and-play synergy with existing frameworks to establish state-of-the-art performance.
Explore related subjects
Keep this discovery
Explore connections, maps & timelines
Younghwan Kil, Joonhyeong Park, Giung Nam, Juho Lee. 2026-11-17. ASCG: Adaptive Spatial Classifier Guidance for Surgical Concept Suppression in Diffusion Models. https://arxiv.science/abs/2611.20625
Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite arXiv Science by Indiciae: BibTeX · RIS.