To overcome the limitations of ground-based telescopes in spatial resolution and imaging quality, recent research has concentrated on image super-resolution (SR) reconstruction. This approach aims to enhance ground-based images by recovering “space-based-like” images with higher spatial resolution and more detailed structural information, without the need for expensive space-based observations. The training of SR models typically relies on large-scale, high-quality paired datasets of ground-based and space-based images. However, the acquisition of such data is highly costly, which significantly limits the widespread application of these models. Based on this, this paper proposes using realistic synthetic galaxy images generated by IllustrisTNG cosmological simulation to replace real space-based images and constructs a training dataset for image SR tasks, TNG-RealSR. To verify the validity of this dataset, we selected five mainstream lightweight SR models for evaluation. The results show that TNG-RealSR exhibits good applicability in the task of galaxy image SR, providing reliable data support for related research. To the best of our knowledge, this work is the first to apply cosmological simulation-generated synthetic galaxy images to the field of galaxy image SR, demonstrating the feasibility of deep learning methods based on synthetic data for astronomical image enhancement tasks. This provides a low-cost and efficient solution for improving data quality in future large-scale surveys. The dataset is available online at https://github.com/jiaweimmiao/TNG-RealSR.