SafeRedir framework unlearns unsafe concepts from image models at inference time

Researchers have developed SafeRedir, a novel framework designed to enhance safety in image generation models by preventing the creation of undesirable content like NSFW imagery or copyrighted styles. This method operates at inference time, redirecting unsafe prompts without altering the original models. SafeRedir utilizes a safety classifier to detect unsafe generation paths and a token-level redirection mechanism to guide prompts toward safe semantic regions, demonstrating effectiveness across various diffusion models. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Introduces a lightweight, inference-time method to improve safety in image generation models without retraining.

RANK_REASON This is a research paper detailing a new method for unlearning in image generation models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• SafeRedir
• Renyang Liu

• paper
• safety