Cross-Domain Image Generation via Graph-Based Structural Modeling

Abstract

This paper addresses the problems of insufficient structural preservation and semantic drift in cross-domain image generation by proposing a cross-domain generation method based on graph-structured modeling. The method constructs node and edge representations of feature maps to explicitly model the semantic regions and structural relationships within images, enabling coordinated optimization of content consistency and style alignment during cross-domain feature transfer. Specifically, the model first uses a shared encoder to extract multi-scale semantic features from both the source and target domains, then builds topological relationships among nodes through a graph construction module. A graph convolution mechanism is applied to propagate and aggregate structured features, allowing the model to learn high-order semantic associations across domains. Finally, a decoder reconstructs the target-domain image. To enhance stability and realism, a joint optimization objective combining adversarial loss, content consistency constraint, and structural preservation term is designed to balance visual quality and structural coherence. Experimental results show that the proposed method outperforms mainstream generative models in MSE, MAE, PSNR, and SSIM, achieving high fidelity and semantic consistency under various cross-domain conditions. This study reveals the intrinsic patterns of cross-domain image generation from a structural modeling perspective and provides an effective new approach for graph-based generative visual modeling.