<p>High-resolution three-dimensional digital rock imaging is fundamental for precise pore-scale and mineralogical characterization. However, computed tomography scans are often limited by coarse voxel resolution, long acquisition times, and imaging noise, which obscure fine structural details and compromise the accuracy of quantitative analysis. The proposed the Physically-Accurate Super-Resolution Transformer (PAST) integrates a dual-path backbone that separates global context modeling from local detail refinement, coupled with a lightweight multi-task prediction head. The global path encodes long-range contextual dependencies at reduced spatial resolution and restores them through learned upsampling, while the local path preserves fine-grained information using hierarchical residual groups composed of Local Window Attention Blocks with a multi-scale windowing strategy. Furthermore, Channel Refinement Blocks adaptively reweight channel responses to enhance structurally discriminative features. To ensure the physical interpretability of the reconstructed data, a physics-aware intensity normalization and reversion pipeline is incorporated, maintaining the quantitative consistency of computed tomography values for subsequent rock property estimation. The super-resolution branch employs a three-dimensional pixel-shuffle operation to generate isotropic high-resolution outputs, and the segmentation branch predicts voxel-wise probability maps that are spatially aligned with the reconstructed volume. Extensive experiments on the DeepRock-super-resolution-3D dataset show that PAST substantially improves reconstruction fidelity, preserves pore connectivity, and enhances segmentation accuracy compared to both convolutional and transformer-based baselines. These results demonstrate the potential of PAST to enable physically consistent and high-precision digital rock analysis.</p>