This paper introduces Matten, a novel latent diffusion model for video generation that integrates Mamba-Attention architecture. Matten employs spatial-temporal attention for local video content modeling and bidirectional Mamba for global video content modeling, achieving competitive performance with current Transformer-based and GAN-based models in benchmark performance, with superior FVD scores and efficiency. The model's complexity directly correlates with video quality, indicating excellent scalability. Recent advancements in diffusion models have shown impressive capabilities in video generation. However, architectural design is crucial for efficient application. Contemporary studies focus on CNN-based U-Net and Transformer-based frameworks, both using attention mechanisms for spatio-temporal dynamics. Spatial attention is widely used in video generation, while local temporal attention is confined to identical positions across frames. Global attention, though effective, is computationally intensive. State space models (SSMs) have gained attention for their ability to handle long sequences. The Mamba model has improved data inference efficiency and model performance by introducing dynamic parameters into SSM structures. Mamba has been successfully extended to vision and multimodal applications. Given the complexity of video data, the authors propose using Mamba to explore spatio-temporal interactions in video content. Matten combines Mamba and Attention mechanisms to capture global temporal relationships and spatial/local temporal relationships. The model variants explore different combinations of Mamba and Attention mechanisms. Comprehensive evaluations show that Matten achieves comparable performance to other models with lower computational and parameter requirements and exhibits strong scalability. The paper also discusses the advantages of Mamba and Attention, introducing a latent diffusion model for video generation with a Mamba-Attention architecture. The model variants investigate the impact of various combinations of Mamba and Attention mechanisms on video generation. The results show that the most effective approach is to use the Mamba module to capture global temporal relationships and the Attention module for spatial and local temporal relationships. The paper presents experimental evaluations of Matten in both unconditional and conditional video generation tasks. Across all test benchmarks, Matten consistently exhibits comparable FVD scores and efficiency with state-of-the-art models. The results indicate that Matten is scalable, evidenced by the direct positive relationship between the model's complexity and the quality of generated samples. The paper also discusses the comparison of Mamba and Attention mechanisms, analyzing their computational efficiency. The results show that Mamba is more computationally efficient than self-attention, particularly for long sequences. For shorter sequences, attention mechanisms are more efficient when computational overhead is manageable. The paper presents experiments on four datasets, showing that Matten consistently generates high-resolution, realistic videos. The results demonstrate that Matten excels in generating high-quality videos on the UCF101 dataset, an area where many other models frequently falter. The quantitative results show that Matten surpasses prior works and matches the performance of methods with image-pretrained weights, demonstrating its superiority in video generation. Additionally, Matten achieves roughly a 25% reduction in flops compared to Latte, the latestThis paper introduces Matten, a novel latent diffusion model for video generation that integrates Mamba-Attention architecture. Matten employs spatial-temporal attention for local video content modeling and bidirectional Mamba for global video content modeling, achieving competitive performance with current Transformer-based and GAN-based models in benchmark performance, with superior FVD scores and efficiency. The model's complexity directly correlates with video quality, indicating excellent scalability. Recent advancements in diffusion models have shown impressive capabilities in video generation. However, architectural design is crucial for efficient application. Contemporary studies focus on CNN-based U-Net and Transformer-based frameworks, both using attention mechanisms for spatio-temporal dynamics. Spatial attention is widely used in video generation, while local temporal attention is confined to identical positions across frames. Global attention, though effective, is computationally intensive. State space models (SSMs) have gained attention for their ability to handle long sequences. The Mamba model has improved data inference efficiency and model performance by introducing dynamic parameters into SSM structures. Mamba has been successfully extended to vision and multimodal applications. Given the complexity of video data, the authors propose using Mamba to explore spatio-temporal interactions in video content. Matten combines Mamba and Attention mechanisms to capture global temporal relationships and spatial/local temporal relationships. The model variants explore different combinations of Mamba and Attention mechanisms. Comprehensive evaluations show that Matten achieves comparable performance to other models with lower computational and parameter requirements and exhibits strong scalability. The paper also discusses the advantages of Mamba and Attention, introducing a latent diffusion model for video generation with a Mamba-Attention architecture. The model variants investigate the impact of various combinations of Mamba and Attention mechanisms on video generation. The results show that the most effective approach is to use the Mamba module to capture global temporal relationships and the Attention module for spatial and local temporal relationships. The paper presents experimental evaluations of Matten in both unconditional and conditional video generation tasks. Across all test benchmarks, Matten consistently exhibits comparable FVD scores and efficiency with state-of-the-art models. The results indicate that Matten is scalable, evidenced by the direct positive relationship between the model's complexity and the quality of generated samples. The paper also discusses the comparison of Mamba and Attention mechanisms, analyzing their computational efficiency. The results show that Mamba is more computationally efficient than self-attention, particularly for long sequences. For shorter sequences, attention mechanisms are more efficient when computational overhead is manageable. The paper presents experiments on four datasets, showing that Matten consistently generates high-resolution, realistic videos. The results demonstrate that Matten excels in generating high-quality videos on the UCF101 dataset, an area where many other models frequently falter. The quantitative results show that Matten surpasses prior works and matches the performance of methods with image-pretrained weights, demonstrating its superiority in video generation. Additionally, Matten achieves roughly a 25% reduction in flops compared to Latte, the latest