<p>Seismic inversion techniques transform seismic amplitude data into quantitative representations of subsurface elastic properties. Stochastic seismic inversion methods are particularly valuable for assessing uncertainty in the spatial distribution of elastic parameters. Iterative geostatistical inversion approaches employ stochastic simulation to generate and adjust models, typically relying on a global variogram to model spatial continuity. However, in complex, non-stationary geological settings, a single variogram fails to capture spatial heterogeneity, leading to suboptimal inversion results. This study presents an innovative 3D geostatistical seismic inversion technique that adaptively updates local variograms by incorporating mismatches between predicted and observed seismic data. By dynamically refining variogram models, the approach addresses the limitations of predefined global variograms. The proposed method, GSI-ALV, is validated using a 3D synthetic non-stationary dataset and benchmarked against Global Stochastic Inversion (GSI), which employs a fixed global variogram. Both methods used identical parameterization, but GSI-ALV achieved a high global correlation coefficient of 0.9 between predicted and observed seismic data, whereas GSI achieved only 0.71. The results demonstrate that dynamically refining variogram models enhances spatial consistency, improves geological realism, and better captures complex subsurface structures compared to conventional methods. Additionally, the proposed approach is specifically designed for 3D seismic inversion and offers significantly lower computational cost than other iterative variogram updating strategies, making it well-suited for large-scale applications.</p>