Abstract: Accurate prediction of bubble spatial distribution in gas-liquid two-phase flow is of vital importance for the design, operation, and safety evaluation of nuclear power equipment. However, existing experimental techniques such as wire-mesh sensors and high-speed imaging systems, as well as numerical simulations based on Euler-Euler or VOF models, often suffer from limited resolution, low temporal-spatial clarity, or high computational cost, making it challenging to obtain detailed bubble distribution information. To address these limitations, this study proposes a neural network model that integrates multi-scale features—dubbed the Fusion model—for the super-resolution reconstruction of bubble images. The proposed model aims to reconstruct high-resolution bubble distribution images (128×128 pixels) from ultra-low-resolution inputs (16×16 pixels), simulating scenarios with extremely coarse measurement or simulation resolution. The Fusion model integrates three subnetworks with skip-connections in a hybrid downsampled skip-connection/multi-scale structure. This model structure allows simultaneous extraction of global structural features and local boundary details, thereby improving the fidelity of reconstructed bubble contours and internal textures. The training dataset was constructed using synthetically generated bubble images with varied bubble sizes (diameters ranging from 10 mm to 30 mm) and spatial distributions, ensuring coverage of representative scenarios in industrial bubble columns. Each sample pair includes a high-resolution image and its corresponding low-resolution counterpart obtained via average pooling. The training process was conducted on an NVIDIA RTX 4090 GPU using the mean squared error (MSE) loss function, with convergence typically achieved within 2.1 h for the Fusion model. Comparative experiments were conducted against two baseline models: a single-bubble enhanced CNN and a multi-scale network without feature integration. In single-bubble cases, the Fusion model achieves a MSE of 0.001, representing a 75% improvement over bicubic interpolation, and a structural similarity index (SSIM) of 0.993 9. In multi-bubble scenarios, the model reduces the MSE by 79.5% (to 0.020 3) and improves SSIM nearly threefold (to 0.899 2), demonstrating superior robustness under varying bubble densities. Additionally, training data sensitivity analysis reveals that the model’s performance stabilizes when the number of training samples exceeds 5 000 pairs, indicating reliable generalization capability. In conclusion, this study validates the effectiveness of the Fusion model in restoring fine-scale bubble features from highly degraded inputs. The method holds strong potential for applications in reactor safety assessment, thermal-hydraulic model validation, and indirect sensor signal enhancement. Future work will focus on improving the model’s physical interpretability by incorporating physics-informed priors and validating the model performance against experimental datasets, paving the way toward real-time, high-resolution bubble monitoring in complex flow environments.