Synthesizing Remote Sensing Images from Land Cover Annotations via Graph Prior Masked Diffusion

Semantic image synthesis (SIS) in remote sensing aims to generate high-fidelity satellite imagery from land use/land cover (LULC) labels, supporting applications such as map updating, data augmentation, and environmental monitoring. However, the existing methods typically focus on pixel-level semantic-to-image translation, neglecting the spatial and semantic relationships among land cover objects, which hinders accurate scene structure modeling. To address this challenge, we propose GMDiT, an enhanced conditional diffusion model that extends the masked DiT architecture with graph-prior modeling. By jointly incorporating relational graph structures and semantic labels, GMDiT explicitly captures the object-level spatial and semantic dependencies, thereby improving the contextual coherence and structural fidelity of the synthesized images. Specifically, to effectively capture inter-object dependencies, we first encode the semantics of each node using CLIP and then employ a simple yet effective graph transformer to model the spatial interactions among nodes. Additionally, we design a scene similarity sampling strategy for the reverse diffusion process, improving contextual alignment while maintaining generative diversity. Experiments on the OpenEarthMap dataset show that GMDiT achieves superior performance in terms of FID and other metrics, demonstrating its effectiveness and robustness in the generation of structured remote sensing images.