With the rapid advancement of artificial intelligence (AI) and deep learning, text-to-face generation has emerged as a powerful tool for bridging natural language and visual content. It not only shows strong capabilities in image synthesis but also presents new opportunities for educational applications, including personalized instruction, history education, and digital literacy. However, current methods still face significant challenges, such as high computational cost in multi-stage models, low training efficiency, and unstable generation quality due to discriminator forgetting. To address these issues, we propose SS-GAN, which includes a skip-stage channel attention excitation module (SSE) and a self-supervised regularized discriminator (SSR). The SSE enhances feature learning by recalibrating channel weights across stages, connecting low- and high-resolution features for better performance. SSR improves discriminator stability by adding decoders trained with self-supervised loss, helping it learn richer facial features. Experiments show that SS-GAN achieves superior generation quality and training stability. Additionally, we explore its potential in educational scenarios, such as personalized virtual teaching assistants, historical figure visualization, and AI ethics education. By aligning technical innovation with educational needs, this work provides new directions and practical support for integrating AI into education. The code is available at https://github.com/caijiWY/SSE-SSR .