Hyperspectral imaging, crucial in remote sensing, provides extensive spectral information at the cost of lower spatial resolution compared to standard color images. Single-image super-resolution, reconstructing high-resolution images from low-resolution inputs, is particularly useful for enhancing hyperspectral images. Due to the unavailability of real low- and high-resolution image pairs, many hyperspectral image super-resolution methods resort to downsampling for training. This leads to suboptimal performance on real-world data due to inherent assumptions in the downsampling process. This paper introduces a novel dataset featuring actual low- and high-resolution hyperspectral image pairs, captured using different lenses and sensors. We train various super-resolution models on this dataset and compare their performance against models trained on artificially downsampled high-resolution images. Our findings reveal that models trained with real image pairs substantially outperform basic bicubic interpolation, whereas those trained with synthetically generated low-resolution images do not, highlighting the importance of using authentic high- and low-resolution images for training.