<p> To address the challenges of three-dimensional reservoir modeling with varying amounts of well data, this study introduces two image processing-inspired generative adversarial networks (GAN): three-dimensional handwritten digit generation GAN (3D-HDG-GAN) and image inpainting GAN (II-GAN). The 3D-HDG-GAN is particularly effective for reservoirs with limited well data, as it leverages three-dimensional convolution operations to learn spatial geological structures. This approach results in an accuracy improvement exceeding 10% compared with the II-GAN when fewer than 40 wells are available. Furthermore, the 3D-HDG-GAN mitigates the spatial information loss encountered during restoration in one-dimensional and two-dimensional HDG-GAN. Conversely, the II-GAN demonstrates superior performance with a larger number of wells, which is enabled by its twin-generator and dual-discriminator architecture. This structure captures detailed spatial information using both standard and dilated convolutions. When the number of wells exceeds 200, II-GAN achieves an accuracy improvement of over 5% compared with the 3D-HDG-GAN, with this advantage increasing as the number of wells grows. Compared to geostatistical approaches and deep learning methods, HDG-GAN and II-GAN efficiently capture planar channel distribution while effectively simulating vertical stacking patterns within sedimentary facies. The influences of the latent vector, computational cost, and data imbalance on 3D-HDG-GAN and II-GAN are also examined. A fully convolutional network architecture proves to be more suitable for 3D-HDG-GAN and II-GAN due to its stability. Meanwhile, II-GAN is more suitable for handling complex and imbalanced reservoir data, which is attributable to its enhanced model capacity driven by higher parameter complexity.</p>