We present Spectral Graph AutoRegressive (SGAR) model, a novel conditional node signal synthesis method for brain networks. Unlike conventional generative models, SGAR employs a coarse-to-fine graph generation strategy in the spectral space: it first predicts low-frequency components that capture the global graph structure and then progressively refines high-frequency details to encode local feature dependencies. SGAR leverages Graph Fourier Transform (GFT) to decompose graph signals in the spectral domain and utilizes a conditional autoregressive transformer to generate spectral components based on disease stage labels. The continuous node signals are subsequently reconstructed via Inverse Graph Fourier Transform (IGFT), preserving the overall network topology. Applied to the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset, our framework effectively addresses the challenges of data scarcity and label imbalance by augmenting brain networks with realistic, structured node features. Experimental results demonstrate that SGAR improves downstream AD classification performance while maintaining the global structure of brain networks.