SafeRedir: A Lightweight Framework for Robust Unlearning in Image Generation Models

Researchers have introduced SafeRedir, a lightweight inference-time framework designed to erase harmful concepts from image generation models (IGMs) without costly retraining. IGMs often memorize undesirable content from training data, such as NSFW imagery and copyrighted artistic styles, posing safety and compliance risks. Post-hoc filtering methods lack robustness and fine-grained semantic control. Existing unlearning methods require retraining, degrade generation quality, or fail against paraphrasing and adversarial attacks. SafeRedir operates via prompt embedding redirection, modifying the model's behavior at inference time to prevent the reproduction of unsafe content while preserving benign generation quality. The framework does not alter the underlying model weights, making it efficient and adaptable. This approach addresses the need for robust, scalable unlearning in real-world deployments of generative AI.