Dynamic generative AI can support digital media learning by generating task-aware tutorials that are adapted to learner behavior, project status, and feedback trajectories. This paper proposes a digital media teaching model DGAI-Tutor. The system integrates software operation logs, project version records, prompt histories, work metadata, and peer feedback from 128 students over 12 weeks to form 6420 annotated learning segments. After data anonymization, timestamp alignment, feature normalization, semantic embedding and time window segmentation, a retrieval augmented large language model is combined with a Transformer learning state encoder and a path adapter. The model generates editorial guidance, concept explanations, and next step tasks for image design, video production, and interactive media projects. Experimental results show that DGAI-Tutor achieves 92.8% tutorial relevance, 89.6% path matching rate and 0.214 feedback MAE, and the 95-quartile delay of three types of tasks is less than 126ms. It is superior to rule retrieval generation and static tutorial recommendation in terms of adaptive guidance and learning experience prediction. The framework can be deployed in the classroom through lightweight caching and controlled generation modules.
