Skip to content

ASCG: Adaptive Spatial Classifier Guidance for Precise Concept Suppression in Diffusion Models

ASCG: Adaptive Spatial Classifier Guidance for Surgical Concept Suppression in Diffusion Models

arXiv: Jailbreaks, Prompt Injection, Poisoning & Defenses· arXiv:2611.20625· Younghwan Kil, Joonhyeong Park, Giung Nam, Juho Lee·Indexed by this site · 2 days ago· Originally publishedDefenses· Its place in the scan record →
AI summary

Researchers propose Adaptive Spatial Classifier Guidance (ASCG), which activates guidance only when harmful content is detected through probabilistic monitoring, and uses Grad-CAM to achieve localized suppression, removing harmful concepts in text-to-image diffusion models while better preserving semantic fidelity than traditional global intervention. The method also introduces a VLM-based evaluation framework to make up for NudeNet's inability to evaluate semantic integrity, and can be plugged into existing frameworks, achieving SOTA performance.

Read originalarxiv.org