Cross-Sequence MRI synthesis enhances diagnostic workflows by reconstructing missing or corrupted contrasts, yet existing unpaired methods often fail to capture both global anatomy and fine textures. In this paper, we introduce FA \(^2\) -Net, a compact dual‑stage framework: Stage I employs frequency‑aware self‑attention within an asymmetric U‑Net to restore macrostructural coherence, and Stage II uses distribution aware residual layer estimation to recover high‑frequency details and quantify uncertainty. To handle unpaired data, FA \(^2\) -Net integrates adversarial and cycle-consistency constraints that promote realistic and anatomically consistent translation across modalities. FA \(^2\) -Net achieves an average PSNR of 25.00 dB and SSIM of 91.30% across six cross-sequence synthesis tasks on the BraTS2021 dataset, outperforming the latest state-of-the-art methods by approximately 3.54 dB in PSNR and 5.47% in SSIM. Similarly, FA \(^2\) -Net attains an average PSNR of 22.14 dB and SSIM of 78.91% across four cross-sequence synthesis tasks on the OpenNEURO dataset, surpassing previous state-of-the-art methods by about 0.93 dB in PSNR and 3.37% in SSIM. Moreover, FA \(^2\) -Net offers a favorable trade-off between computational complexity and accuracy, requiring moderate training and inference times with reasonable memory consumption compared to competing methods. Our code repository will be available at FA-Net .