Thermal infrared (TIR) images play a crucial role in applications such as object detection and autonomous driving. However, unlike high-resolution visible light cameras, which are now widely available, capturing high-resolution TIR images typically requires costly specialized equipment. In this paper, we propose MAGNet, a novel model designed to synthesize high-resolution TIR images from low-resolution TIR images and high-resolution visible light images. Our approach first employs a CNN to extract multi-scale features from both input images. These features are then fused using a Swin Transformer, producing a series of hierarchical feature maps. Finally, a sequence of convolutional and upsampling layers decodes these hierarchical representations into a high-resolution TIR image. Unlike other methods, MAGNet combines multi-level CNN and Swin Transformer for image fusion and reconstruction, which improves the SR results, especially in complicated scenarios. We evaluate MAGNet on the VGTSR dataset, where it outperforms state-of-the-art methods. Furthermore, its symmetric architecture allows for potential extension to cross-modal image fusion, real-time thermal enhancement, and unaligned SR tasks.