Dynamic Attention-Guided Video Generation from Text with Multi-Scale Synthesis and LoRA Optimization

This paper introduces DAG-VSNet, a new architecture for high-quality text-to-video generation. The model uses a Dynamic Text Encoder (DTE), Multi-Scale Video Generator (MS-VG), Dynamic Attention Map (DAM), and LoRA-Enhanced Video Generator (LVG). The DTE captures temporal patterns, and the DAM helps focus on important text features. The MS-VG improves video details at different scales, and the LVG fine-tunes pre-trained weights with less computation. Experiments show DAG-VSNet works better than baseline models in metrics like Inception Score (IS), Fréchet Inception Distance (FID), Temporal Consistency Score (TCS), and Text-Video Alignment Score (TVAS). Ablation studies show that LoRA, dynamic attention, and multi-scale methods improve performance. This work provides a useful framework for future research in text-driven video generation.