This paper addresses the graph alignment problem of identifying corresponding nodes in multiple graphs. Existing unsupervised learning-based methods embed node features into latent representations to perform cross-graph comparisons without ground truth correspondence. However, they suffer from reduced node independence due to excessive smoothing of GNN-based embeddings and latent space inconsistency caused by structural noise, feature heterogeneity, and learning instability. In this paper, we propose a novel graph alignment framework that enhances node independence and enhances geometric consistency across latent spaces. A double-pass encoder combining low- and high-pass spectral filters is used to generate structure-aware, high-dimensionally discriminative embeddings. A geometry-aware feature map module is integrated to learn bijective and equidistant transformations between graph embeddings, ensuring consistent geometric relationships between different representations. Comprehensive evaluations on vision-language benchmarks using various pretrained models demonstrate that the proposed framework generalizes beyond the graph domain, enabling alignment of vision and language representations using unsupervised learning.
