The increasing deployment of large generative models has heightened concerns over security and privacy, particularly regarding the generation of inappropriate content such as violent, explicit, or sensitive images, as well as the potential for creating fake images that spread misinformation and cause social problems. In this work, we propose Guided Safe Diffusion (GSD), an inference-time method specifically designed for diffusion models to prevent the generation of images with undesirable content as defined in a prohibited content list. Our method integrates safety guidance during the denoising steps of the model’s inference process, modifying the predicted noise to steer the generation process away from unwanted content. This approach allows the model to accept both an input image and a text description, facilitating controlled image generation. Unlike previous methods, our technique does not necessitate retraining or fine-tuning of the model. We conduct qualitative and quantitative experiments to assess the effectiveness of our method, demonstrating that GSD can remove the unwanted content while preserving unrelated content. The results validate our method’s ability to mitigate risks while maintaining the generative utility of diffusion models.