This paper introduces a Feature Alignment Generative Adversarial Network (FAGAN) for multi-scale fusion modeling of digital core images, aiming to address the trade-off between field of view (FOV) and resolution by combining 2D high-resolution images with small FOV (2D HRI) and 3D low-resolution images with large FOV (3D LRI). Given the dimensional and representational differences between 2D and 3D image features, the generator in our model employs dual-stream networks to extract semantic features from 3D LRI and 2D HRI, respectively. A feature reconstruction module is designed to transform the 2D image feature \(F_{\text {2D}}\) into a 3D feature representation \(F_{\text {2D} \rightarrow \text {3D}}\) , enabling effective feature fusion in the 3D feature space between 2D HRI and 3D LRI. To ensure the semantic consistency of the reconstructed features, a feature space alignment loss function (FSAloss) is introduced to constrain the semantic feature distribution of \(F_{\text {2D} \rightarrow \text {3D}}\) and \(F_{\text {2D}}\) . Additionally, a feature alignment module is designed to further ensure the correct alignment of \(F_{\text {2D} \rightarrow \text {3D}}\) with the semantic features \(F_{\text {3D}}\) of the 3D LRI, thereby effectively fusing the features of both. Visualization results verify that our proposed model can generate structures that not only conform to the spatial geometric features of the 3D LRI but also possess the fine detail features of the 2D HRI. Finally, the effectiveness of the reconstructed results is further validated through the statistical parameters (such as aperture distribution, shape factor distribution, etc.) of the reconstructed structures and numerical analysis (such as permeability, etc.).