DiSCO: Black-Box Defense Against NSFW Text-to-Image Generation

DiSCO provides a zero-shot, plug-and-play defense against text-to-image models by optimizing prompts through distribution-guided contrastive methods. It specifically targets benign adversarial prompts that trigger harmful outputs in proprietary models without requiring access to model weights.