A spatially adaptive Transformer-based generation network (SAT-StyTR) was proposed to address the issues of existing face photo-to-sketch synthesis methods, such as structural distortion, blurred textures, and artifact residue. A dual-modal feature disentanglement module (DFDM) was constructed in the encoder. The parallel branches of face content and sketch style feature extraction were employed in DFDM to achieve disentangled representation of cross-modal features. A row-column vector dynamic fusion module (VDFM) was designed, and a bidirectional gating mechanism was introduced in the VDFM to enable adaptive feature fusion of row-column vectors. Then the capability of local detail retention was significantly improved. A multi-scale perceptive discriminator (MSPD) was incorporated to optimize the realism of generated sketch textures at multiple semantic levels by leveraging the hierarchical gradient supervision mechanism. The experimental results on the CUHK, AR, XM2VTS and CUFSF datasets demonstrated that the proposed method outperformed existing methods such as Mamba-ST on SSIM, PSNR, MSE, LPIPS, and FID metrics, with its FID values improved by 19.3%, 1.0%, 15.5%, and 10.0% over those of the second-best method on the four datasets, respectively. Visual comparative analysis further validated that superior structural consistency and detail restoration were achieved in critical regions, such as facial contours.