Small datasets can hinder the training of robust machine learning models. What if we could create more data? Researchers have been exploring synthetic data generation, with techniques ranging from simple statistical methods to complex large language models (LLMs). But a new player has entered the game: TAEGAN. This innovative approach combines the power of generative adversarial networks (GANs) with the ingenuity of masked autoencoders (MAEs). Traditional GANs struggle with small datasets and complex feature relationships. LLMs, while powerful, can be computationally expensive and prone to overfitting, not to mention memorizing, small data. TAEGAN offers a solution by pre-training a MAE to reconstruct masked data, forcing it to learn complex feature interdependencies and avoid the computational overhead of LLMs. This pre-trained MAE then becomes the generator in a GAN framework, producing realistic synthetic tabular data. What sets TAEGAN apart is its clever use of 'hint vectors' and a 'multivariate training-by-sampling' method. These innovations allow the model to generate diverse and representative synthetic data that enhances model performance on small datasets. In tests across various small datasets, TAEGAN consistently improved the performance of downstream machine learning models. TAEGAN outperformed all other tested deep learning models when measured by synthetic data quality metrics and achieved the best data augmentation performance on the majority of the datasets. Its efficient design allows it to avoid the computational bottleneck of larger models, opening doors to new possibilities in data augmentation.