Accurate prediction of the spatial distribution of gas–liquid two-phase flow bubbles is crucial in the design and operation of nuclear energy equipment, however, high-precision experimental and numerical simulation data are usually difficult to obtain, with high cost and long computation time. Traditional experimental methods and numerical simulation techniques can only provide low-resolution bubble distribution information, which is difficult to meet the needs of high-precision gas–liquid two-phase flow research. For this reason, a neural network-based multiscale super-resolution reconstruction method is proposed in this paper, which realizes high-resolution reconstruction of low-resolution bubble distribution data by training the model. The results show that in the fourfold pooling test case, the MSE of the multiscale model is lower than 0.2%, and the SSIM can reach more than 98%. In the more difficult eightfold pooling test case, the MSE of the multiscale model is reduced by up to 91.9% compared with bicubic interpolation, and the SSIM is improved by up to 282%, showing excellent reconstruction results. This paper also explores the effect of the number of training sets on the model performance, and finds that there is a smooth turning point in the number of training sets, and the effect of further increasing the dataset size on the model performance improvement gradually diminishes. The method in this paper demonstrates the potential of super-resolution reconstruction in the field of bubble distribution refinement and provides a new research direction for the application of super-resolution reconstruction in gas–liquid two-phase flow. Future research can further introduce physical a priori knowledge to enhance the model performance and promote the application of super-resolution technology in complex gas–liquid two-phase flow fields.