The current field of digital media art faces two core problems: (1) Video analysis and semantic understanding suffer from problems such as insufficient extraction of spatiotemporal features and insufficient multimodal data fusion capabilities; (2) Traditional video content generation relies on manual creation, which is inefficient and difficult to achieve dynamic adaptation of complex scenes. This study uses an improved 3D Swin Transformer to construct a spatiotemporal attention mechanism, and extracts multi-scale spatiotemporal features of video frame sequences through a hierarchical pyramid structure. A multimodal feature alignment module based on contrastive learning is designed, and the cross-modal embedding space of the CLIP pre-trained model is used. In the creation stage, based on the spatiotemporal features extracted in the analysis stage, a two-way diffusion generation architecture is proposed: the content branch adopts a latent diffusion model; the style branch injects style statistics through the AdaIN layer to control the artistic expression of the generated content. Comparative experiments show that the maximum creation time of the model is only 6.59 min, which is much lower than the 59.4 min of traditional manual creation. In terms of style diversity index, the model also maintains a leading position, proving its efficiency and diversity in dealing with complex content generation tasks. This paper provides new possibilities for content creation in the field of digital media art and promotes the deep integration of technology and art.