<p>Coral reef limestone (CRL), a biogenic marine sedimentary rock, is characterized by low density, high friability, strong anisotropy, and significant inter-sample variability. These physical and mechanical properties, governed by its complex pore structure, pose substantial challenges for sampling and numerical simulations. This study presents a comprehensive generative adversarial network (GAN)-based framework for reconstructing 2D slices and 3D digital rock models of CRL, incorporating five distinct approaches: stochastic reconstruction, porosity-controlled reconstruction, structure-controlled reconstruction, coupled porosity-structure reconstruction, and layer-sequence reconstruction. The study further analyzes the influence of sampling methods on structure-controlled reconstruction performance. The generated images and models are rigorously evaluated using image-based metrics to assess reconstruction fidelity, morphological diversity, and conditional control accuracy. Results indicate that the proposed GAN framework outperforms conventional multiple-point statistics (MPS) algorithms in both reconstruction realism and conditional control precision. The methodology enables (<i>i</i>) the generation of numerous statistically representative CRL samples for numerical simulations and (<i>ii</i>) the production of samples with varying porosities while preserving structural morphology, thereby supporting parameter-controllable numerical studies. Additionally, this framework provides a methodological foundation for reconstructing other rock types and predicting geological image data.</p>