Acquiring submarine wake flow data is both expensive and challenging. Traditional generative models struggle with small sample sizes and multi-scale feature capture, making it difficult to preserve wake image characteristics and generate high-quality results from limited data. We introduce MAPS-GAN (Multi-scale Attention Progressive Self-attended GAN), incorporating a Self-Attention Block (SAB) for balanced feature extraction and representation, along with a progressive generation strategy and lightweight structure for effective multi-scale information capture. The model includes a Multi-scale Skip Excitation (MSE) module to enhance multi-scale feature perception, while its generation framework accelerates training. Experimental results demonstrate that MAPS-GAN effectively overcomes data scarcity and outperforms traditional methods in image quality and computational efficiency, achieving SSIM, PSNR, and FID scores of 0.812, 25.21, and 6.70, respectively.