Abstract: Spectral modeling faces challenges such as high data acquisition costs, limited data sources, and difficulties in data labeling. These issues give rise to the small sample issue, which often leads to insufficient representativeness of training data and difficulties in feature extraction, thereby undermining model stability and generalization ability. Generative Adversarial Network (GAN), as a prominent virtual sample generation technique, offers an effective solution to the small sample issue in spectral analysis. However, the conventional GAN suffers from inherent limitations such as training instability and mode collapse, making it insufficient to meet the demands of spectral modeling. This paper systematically reviewed the theoretical foundation and development of GAN technology, analyzed its limitations, and proposed targeted improvement techniques as well as extension methods for multiple application scenarios. Furthermore, it explored the specific applications of these improvement and extension methods in enhancing the stability and generalization of spectral models. This study aims to provide both theoretical and practical references for building efficient spectral analysis models, thereby promoting the innovation and advancement of spectral analysis technology.