The failure mechanism of internal seepage flow in geotechnical bank slopes is complex. Investigating the microscale characteristics of pore structure and seepage flow behavior forms the scientific foundation for preventing slope collapse. This study focuses on Loose Gravel Soil from the Heishanxia reservoir area of the Yellow River. Multi-view stereo modeling technology is employed to scan and reconstruct coarse aggregate particles above d₁₀ (0.14 mm), d₂₀ (0.20 mm), and d₃₀ (0.28 mm) from the gradation curve into clumps through multi-angle photogrammetric reconstruction, while finer particles below these thresholds are modeled as spherical grains. Three virtual samples (B-10, B-20, B-30) with varying proportions of spherical particles were developed in PFC 5.0 using precompaction methods. The rationality of the modeling approach is verified by comparing the calculated permeability with experimentally measured values. To further validate the methodological efficacy, a variable hydraulic gradient soil column experiment replicating reservoir operational conditions was implemented, followed by a comparative analysis of seepage characteristics through Stokes flow solver simulations. Following the selection of representative elementary volumes (REV) using ninepoint sampling, quantitative analyses of pore-space topology, including coordination number, pore size, pore morphology, and fractal dimension, are conducted. Three-dimensional single-phase flow simulations are performed along the X/Y/Z directions using pore network modeling (PNM), and the seepage characteristics were analyzed. The proposed modeling methodology establishes a foundational framework for the future prevention and control of seepage flow-induced instability in the Yellow River Stone Forest sandy conglomerate bank slopes.