📄 中文摘要
安全机制在扩散和流模型中的发展分为两条不同的路径。在机器人规划中，通过控制障碍函数在每个去噪步骤中显式施加几何约束，引导生成轨迹远离障碍物。同时，最近的数据驱动负引导方法显示出抑制有害内容并促进生成样本多样性的能力。然而，这些方法依赖于启发式规则，并未明确说明何时需要安全引导。提出了一种统一的概率框架，利用最大均值差异（MMD）潜力用于图像生成任务，将受保护的扩散和安全去噪器重新表述为特定实例。
安全引导流（SGF）：安全生成中负引导的统一框架
📄 English Summary
Safety-Guided Flow (SGF): A Unified Framework for Negative Guidance in Safe Generation
This research presents a unified probabilistic framework utilizing Maximum Mean Discrepancy (MMD) potential for image generation tasks, effectively recasting Shielded Diffusion and Safe Denoiser as specific instances. Safety mechanisms for diffusion and flow models have evolved along two distinct paths: in robot planning, control barrier functions guide generative trajectories away from obstacles at each denoising step by imposing geometric constraints. Concurrently, recent data-driven negative guidance approaches have demonstrated the ability to suppress harmful content and enhance diversity in generated samples. However, these methods rely on heuristics without clearly defining when safety guidance is necessary.