GHS-VDG:Graph and Hybrid Spatio-Temporal Attention for Video Diffusion Generation

Video generation is a core task in computer vision, its key challenge is the effective modeling of the temporal and spatial dependencies between video frames. This paper proposes a diffusion model-based video generation framework, GHS-VDG, which introduces the Video Frame Temporal Graph Aggregator (VFTGA) and Hybrid Spatio-Temporal Attention (HSTA) modules to capture complex spatiotemporal dependencies between video frames, thereby enhancing the quality and accuracy of predicted frames. Specifically, in the VFTGA module, we innovatively represent video frames as graph nodes and construct edge connections according to their temporal order, which effectively enhances the capability of capturing temporal information. The HSTA module combines temporal and spatial attention mechanisms to optimize the learning of spatiotemporal features. Experimental results demonstrate that GHS-VDG performs excellently on datasets such as SMMNIST, BAIR, KTH, and UCF101, validating its effectiveness in video generation tasks.