Generative Adversarial Networks (GANs) can synthesise high-fidelity and novel images with a random noise vector and have been widely used in various scenarios, such as image style transfer, image translation, etc. High-quality generation mostly relies on a substantial set of samples for training. However, obtaining a large volume of data can be challenging in some situations due to copyright restrictions, privacy concerns, data scarcity, etc. To overcome this problem, we designed a novel GANs model inspired by the hide-and-seek game played in childhood for generations with limited target data. Specifically, we developed this concept into the DCGANs models by introducing an additional shelter dataset and a shelter discriminator. Instead of the one-to-one competitive game (generator VS discriminator) in the standard DCGANs, we employed a one-to-two (generator VS shelter discriminator + hider discriminator) learning strategy. The novel Hide-and-Seek GANs can generate semantic images from limited data and achieve style mixing. Our experiments demonstrated that the solution was effective and allowed us to discover more interesting results. Remarkably, the work provides a compelling example of AI innovation education for teenagers.