DiSCO: Defending text-to-image generation through distribution-guided contrastive prompt optimization | ResearchPod