Professor Kim Mingyu of the Department of Artificial Intelligence at the College of Software Convergence, Kookmin University (President Jeong Seung Ryul), has conducted research proposing “Safety-Guided Flow (SGF),” an integrated framework for safe content generation based on diffusion models and flow matching—two leading techniques in image and video generation AI.

This research was supported by the Institute for Information & Communications Technology Planning & Evaluation (IITP) and CIFAR in Canada. Professor Kim Mingyu served as the first author, with Professors Kim Young-heon and Park Mi-jeong from the University of British Columbia (UBC) in Canada participating as co-authors.

The research results were accepted for an oral presentation at the 『International Conference on Learning Representations 2026』, one of the world’s top three conferences in the field of artificial intelligence (ranked 8th in all science and engineering fields based on Google Scholar H-index).

According to the study, it demonstrated that existing representative safety generation methods—Shielded Diffusion and Safe Denoiser—are special cases of the Maximum Mean Discrepancy (MMD) potential, thereby unifying the previously fragmented field of safety generation research into a single integrated framework. Furthermore, by applying Control Barrier Function theory, the study demonstrated the existence of a “critical time window” during which guidance must be applied strongly at the beginning of the denoising process and then gradually reduced. Models utilizing SGF demonstrated superior performance compared to existing methods across various safe generation scenarios, including harmful content defense, prevention of training data memorization, and copyright protection, confirming its potential as a core foundational technology for the safe practical application of generative AI.

Professor Kim Mingyu of Kookmin University stated, “This study theoretically proves that existing data-based safe generation models are special cases of the probability-based Maximum Mean Discrepancy potential gradient, thereby presenting a new analytical framework that allows us to understand the previously fragmented field of safe generation research from a unified perspective.” He added, “In the future, diffusion models and flow-matching models are expected to be utilized as core foundational technologies to ensure safety as they are deployed in high-risk domains such as autonomous driving, healthcare, and content generation.”