UniNDM: A Unified Noise-Driven Sexual Content Detection and Mitigation Framework for Text-to-Image Generation

UniNDM proposes a unified noise-driven framework that detects and mitigates inappropriate content generated by text-to-image diffusion models under implicitly suggestive prompts. The framework exploits the separability of early diffusion predicted noise between normal and pornographic content to build a lightweight detector with near-zero overhead, and uses an LLM to dynamically generate context-relevant negative prompts and suppress attention concentration on explicit tokens, while also extending to DiT architectures. On natural and adversarial datasets for U-Net and DiT models, it outperforms methods such as SLD, UCE, and Safree.