SafeRedir: Prompt Embedding Redirection for Robust Unlearning in Image Generation Models

Researchers introduce SafeRedir, an inference-time framework that safely redirects unsafe prompts in image generation models by rerouting them toward benign semantic regions without modifying underlying model weights. The lightweight approach uses token-level embedding interventions to mitigate generation of NSFW content and copyrighted styles while maintaining image quality and resisting adversarial attacks.

SafeRedir addresses a critical vulnerability in modern image generation systems: their tendency to reproduce memorized unsafe content despite post-deployment filtering attempts. Traditional unlearning approaches require expensive model retraining, degrade output quality, or crumble under prompt manipulation attacks. This research proposes an elegant alternative by intervening at the embedding level during inference, effectively creating safety guardrails without architectural changes.

The framework's two-component architecture—a safety classifier identifying unsafe trajectories and a delta generator for semantic redirection—represents a meaningful advance in AI safety engineering. By operating at token granularity rather than image-level filtering, SafeRedir achieves precision that coarser methods cannot match. The approach's plug-and-play compatibility across different diffusion model backbones significantly increases its practical deployment potential.

For organizations deploying image generation models in regulated industries, SafeRedir offers meaningful risk reduction without performance sacrifices. Its demonstrated robustness against adversarial prompts addresses a persistent gap in existing safety mechanisms. The framework's ability to preserve benign generation quality while suppressing harmful outputs creates a more viable path to responsible AI deployment than current alternatives.

The research validates effectiveness across multiple unlearning tasks and demonstrates generalization to existing unlearned models. However, the real-world impact depends on adoption by model providers and integration into production pipelines. The open-sourced codebase could accelerate industry adoption, particularly among developers prioritizing compliance and safety in creative AI applications.

• →SafeRedir redirects unsafe prompts at the embedding level without modifying underlying image generation models.
• →The framework demonstrates strong robustness against prompt paraphrasing and adversarial attacks compared to existing unlearning methods.
• →Token-level interventions preserve image quality and semantic coherence better than traditional retraining-based approaches.
• →The solution generalizes across multiple diffusion model architectures with plug-and-play compatibility.
• →Inference-time deployment eliminates computational costs and quality degradation associated with model retraining.