Keep It CALM: Analyzing the Limits of Global Unsafety in Text-to-Image Generation

Training-free safeguards for text-to-image generation often rely on a reusable safety signal, such as an unsafe direction or global toxic subspace, applied broadly across prompts.