<p>This study presents a Multi-Grid Descriptor-based workflow for three-dimensional reconstruction of rock microstructures from limited two-dimensional slice images. The framework integrates multi-scale feature extraction, descriptor matching, Structure from Motion, and automated point-cloud processing. Five reconstruction strategies were systematically compared: single-slice reconstruction with and without multi-grid descriptors, ten-slice reconstruction, twenty-slice reconstruction, and three-view reconstruction with multi-grid descriptors. Reconstruction quality was evaluated against the micro-CT reference using RMSE, porosity deviation, autocorrelation, variogram analysis, and morphological descriptors including Kozeny-Carman permeability, Euler density, box-counting fractal dimension, local porosity heterogeneity, chord anisotropy, and pore-cluster connectivity. Results show that slice-dependent stochastic reconstruction strategies exhibit metric-specific advantages rather than a universal optimum. The ten-slice method achieved the lowest RMSE, indicating better voxel-level numerical accuracy. The twenty-slice method better approximated several morphology- and transport-related descriptors, whereas the three-view multi-grid method better preserved selected topological and heterogeneity characteristics. All reconstructions underestimated porosity and failed to recover percolating pore clusters, indicating limited recovery of long-range connectivity. These findings suggest that reconstruction strategies should be selected according to downstream analysis goals rather than a single accuracy metric.</p>