The rapid advancement of diffusion models (DMs) has raised significant concerns regarding their safe use, as these models can inadvertently generate inappropriate, not-safe-for-work (NSFW) content, copyrighted material, or private data. In response, recent research on safe diffusion models has explored techniques such applying text-based negative prompts, or retraining models to suppress certain features. While these approaches have shown promise, they often come with limitations, including reliance on external classifiers, loss of generation quality, or substantial computational costs due to fine-tuning.

In this talk, I will first provide an overview of existing methods for enhancing the safety of diffusion models and discuss their respective strengths and limitations. Building on this foundation, I will introduce our recent work, which takes a fundamentally different approach to ensuring safe image generation. Instead of modifying the model itself, we propose a training-free safe denoiser that directly adjusts the sampling trajectory to avoid unsafe or restricted regions of the data distribution. By leveraging a negation set—such as unsafe images, copyrighted data, or private information—we derive a denoiser that ensures final samples remain outside these areas without requiring additional training or fine-tuning. Our method is effective across various text-conditional, class-conditional, and unconditional generation settings, demonstrating the potential of training-free strategies for safer diffusion model deployment.

I’m a Postdoctoral Research Fellow at The University of British Columbia (UBC) working with Prof. Mi-Jung Park. I received a Ph.D. in Artificial Intelligence from KAIST under the supervision of Prof. Se-Young Yun. Prior to beginning my PhD course, I worked as a research engineer at Samsung Heavy Industries. I received an M.S. in Ocean system engineering from KAIST under the supervision of Prof. Hyun Chung and B.S in Industrial engineering at Konkuk University. My research interests focus on probabilistic machine learning, bayesian approach, neural fields, multimodal learning, few-shot learning, generative models, bayesian models and their applications.