<p>In recent years, stacked a deeper model can get a better super-resolution result, but a large deep model is difficult to train. In this paper, to address the difficulty of training a huge deep image super-resolution model, we propose an efficient progressive training framework with granularity cross for image super-resolution (EPTGC). Splitting the model and combining the use of multiple images with different granularities, not only reduces the difficulty of training the model but also helps the model to learn different granular features, and enhance the ability of the model to recover the image edge information. EPTGC is a plug-and-play approach and can be applied to most image super-resolution models. In the experiments, we apply EPTGC to 8 different models, including the convolutional neural network (CNN) based models and transformer-based models. The results show that using EPTGC improves their ability to recover the edge information of the image and improves their results on the 4 benchmark datasets by a maximum of 0.44 DB on PSNR.</p>