As a popular research field of computer vision, super-resolution is currently widely studied. In the past, the size of the training set required for super-resolution work was too large. A large training set would cause more resource requirements, and at the same time, the time overheads of data transmission would also increase. Moreover, in super-resolution work, the relationship between the complexity of the image and the model structure is usually not considered, and images are recovered in same depth. This method often cannot meet the SR-reconstruction needs of all images. This paper proposes a new training and reconstruction framework based on multiple models. The framework prunes the training set according to the complexity of the images in the training set, which significantly reduces the size of the training set. At the same time, the framework can select the specific depth according to the image features of the images to recover the images, which helps to improve the SR-reconstruction effect. After testing different models, our framework can reduce the amount of training data by 41.9% and reduce the average training time from 2935 min to 2836 min. At the same time, our framework can improve the average SR-reconstruction effect of 65.7% images, optimize the average perceptual index from 3.1607 to 3.0867, and optimize the average SR-reconstruction time from 101.7 s to 66.7 s.