This paper presents a unified variational framework for high-fidelity dense depth reconstruction from passive imaging systems. We formulate depth estimation as a global energy minimization problem, integrating data fidelity terms derived from stereo correspondence and focus measures with a Total Variation (TV) regularizer. The proposed Alternating Direction Method of Multipliers (ADMM) solver efficiently handles the non-smooth convex optimization, while a Random Walk with Restart refinement post-processes confidence. Extensive experiments on high-precision synthetic facial datasets and real-world automotive street scenes validate the method's robustness across modalities. Quantitative results demonstrate superior convergence and lower RMSE compared to traditional gradient-based methods, effectively mitigating noise while preserving critical depth discontinuities. The framework establishes a mathematical foundation for modality-agnostic depth estimation.