As educational digitalization accelerates, the need for intelligent technologies in teaching and learning is becoming increasingly pressing. This study introduces a deep hybrid recommendation framework, VAE-GAN-DCR, built on variational autoencoders (VAE) and generative adversarial networks (GAN), and examines how generative AI can contribute to smart-classroom practice. Methodologically, the model integrates the VAE decoder with the GAN generator, extends standard VAE design by incorporating item-feature–dependent prior distributions, and reduces reconstruction error by leveraging feature-transfer signals from the GAN discriminator to improve educational resource recommendation accuracy. In addition, teaching outcomes at Zhanjiang Early Childhood Teacher Training College are assessed using the Williams Creative Tendency Measurement Scale. Experimental results show that VAE-GAN-DCR achieves strong performance across three datasets; on MovieLens-1M, Recall@20 increases by 12.15% and NDCG@100 by 12.94%. Classroom application results further indicate that the experimental group outperforms the control group in creative-thinking activities and creative tendency, with the overall creative tendency score reaching 2.61. The findings suggest that generative AI can both enhance the precision of educational resource recommendations and foster students’ creativity, providing robust support for smart-classroom development.