Summary Seismic surveys are often affected by irregularity and coarse sampling resulting from limitations at the acquisition such as obstacles and environmental constraints. Thus, various processing algorithms have been developed to reconstruct these missing data, including interpolation based on transforms and deep learning-based reconstruction methods, where the latter allows the extraction of more complex structures from tons of data. Alternatively, deep-image-prior (DIP) methods mitigate the use of extensive training information by considering only the data under observation, showing promising results in recovering seismic traces. This work proposes a DIP-based convolutional neural network for recovering missing sources in a three-dimensional (3D) seismic survey which exploits as priors the data acquired and the structure of the 3D network. Additionally, considering the correlation across the source domain, an operator is employed for vertically shifting the set of traces to have the maximum data correlation in the network. Numerical experiments show that the proposed network recovers the complete set of traces in each shot-gather, where the information and events are preserved even though the source is not acquired.