Precipitation detection using infrared (IR) brightness temperature (BT) temporal flux data is a challenging problem. Other sensors, such as microwave (MW), have reliable and more robust predictive performance, but lack land coverage and temporal availability. IR-BT provides high-frequency data (from half an hour to 10 min) at very low resolution (4 km). However, automatic precipitation detection frameworks should face the simple nature of this variable on the one hand, and the very low number of rain events occurring in nature on the other hand. This paper addresses this challenge by proposing a conditional GAN framework using recurrent neural networks, which transforms the unbalanced problem into a small (short) pattern detection algorithm. Several tests allow the identification of robust architectures and useful loss functions that enable promising results, minimize false alarms, and improve the overlap of positive events.