Building Information Models of cultural heritage structures present a fundamental challenge for real-time Virtual Reality applications: while heritage documentation demands geometric fidelity at Level of Detail 350–400, interactive VR requires strict performance constraints exceeding 60 frames per second. This study presents a hybrid optimization framework combining Quadric Error Metrics edge collapse with adaptive decimation parameters and occlusion culling to bridge this gap. Applied to St. Petersburg Polytechnic University’s neoclassical campus established in 1902, the method achieved 7.5× geometric compression for the main building, reducing from 71.8 to 9.6 million polygons while maintaining Hausdorff distance below 5 cm and sustaining 52 ± 3 frames per second on Oculus Quest 2. Compared to baseline methods, Revit’s native Level of Detail at 40% reduction and manual retopology achieving 60% reduction over 80 h—the automated approach achieved 86% reduction in 4 h with preserved ultraviolet coordinates, enabling photorealistic texture preservation. Validation employed geometric metrics including Hausdorff distance and Structural Similarity Index, production metrics encompassing frames per second and draw calls, and user assessment involving 25 trained evaluators. Results confirm that mathematically grounded mesh decimation, combined with semantic geometry filtering, provides an efficient pathway for integrating high-fidelity heritage documentation into accessible VR ecosystems.