With the rapid development of technology, its applications in various industries have become increasingly widespread, particularly in the field of security detection. X-ray security screening, as a critical non-contact detection technology, plays an essential role in security inspections at public places such as airports, train stations, and borders. However, the application of advanced technology in X-ray security screening faces a significant challenge: the lack of real and large-scale X-ray security images for model training. Traditional methods of acquiring X-ray security images are time-consuming and labor-intensive, constrained by safety and ethical requirements. To address this issue, we propose a novel method utilizing pre-trained controllable diffusion models to synthesize realistic X-ray images for training purposes. Our approach incorporates a Hypercomplex Spatial Channel Attention (HSCA) module with hypercomplex attention and quaternion computations within the encoder of a Variational Autoencoder (VAE). This innovative attention module enhances the synthesis effect by improving foreground detail preservation and attribute accuracy. Extensive experiments demonstrate that our method effectively generates high-quality X-ray images that closely align with real-world scenarios, significantly enhancing the performance of X-ray security screening models.