This paper provides a comprehensive survey of recent advancements in long video generation, focusing on two key paradigms: *divide and conquer* and *temporal autoregressive*. The authors delve into the common models used in each paradigm, including network design and conditioning techniques. They also offer a detailed overview and classification of datasets and evaluation metrics crucial for advancing long video generation research. The paper concludes with a summary of existing studies, discussing emerging challenges and future directions in this dynamic field. Key contributions include: 1. **Definition of Long Videos**: The paper proposes a clear definition of long videos as those exceeding 10 seconds or containing more than 100 frames. 2. **Model Paradigms**: - **Divide and Conquer**: This paradigm involves generating keyframes first and then filling in the intervening frames to create a cohesive long video. - **Temporal Autoregressive**: This approach generates short video segments based on preceding conditions, ensuring fluid transitions between clips. 3. **Control Signals**: Text and image prompts are used to guide the generation process, ensuring content, style, and thematic relevance. 4. **Challenges and Solutions**: - **Temporal-Spatial Consistency**: Techniques like adding temporal attention layers and using temporal discriminators enhance temporal-spatial consistency. - **Content Continuity**: Methods such as training directly on long videos and using autoregressive strategies improve content continuity. - **Diversity**: Efforts to generate high-resolution, variable-size, and diverse content are discussed. 5. **Resource Management**: - **Data Compression**: Techniques like T-KLVAE and VAE encoders reduce data dimensionality and computational complexity. - **Lightweight Model Design**: Mask modeling and bidirectional transformers enhance efficiency. - **Training Strategies**: Pre-training on large-scale text-to-image datasets and supplementary dataset methods address resource constraints. 6. **Future Directions**: - **Expansion of Data Resources**: Methods to enrich long video datasets. - **Unified Generation Approaches**: Integrating the strengths of both paradigms. - **Flexible Length and Aspect Ratio**: Enabling generation of videos with variable dimensions. - **Super-Long Videos**: Addressing challenges in generating videos longer than one hour. - **Enhanced Controllability and Real-world Simulation**: Improving understanding and control over generation models. The paper aims to serve as a valuable reference for researchers and practitioners in the field of long video generation, highlighting both current advancements and future research directions.This paper provides a comprehensive survey of recent advancements in long video generation, focusing on two key paradigms: *divide and conquer* and *temporal autoregressive*. The authors delve into the common models used in each paradigm, including network design and conditioning techniques. They also offer a detailed overview and classification of datasets and evaluation metrics crucial for advancing long video generation research. The paper concludes with a summary of existing studies, discussing emerging challenges and future directions in this dynamic field. Key contributions include: 1. **Definition of Long Videos**: The paper proposes a clear definition of long videos as those exceeding 10 seconds or containing more than 100 frames. 2. **Model Paradigms**: - **Divide and Conquer**: This paradigm involves generating keyframes first and then filling in the intervening frames to create a cohesive long video. - **Temporal Autoregressive**: This approach generates short video segments based on preceding conditions, ensuring fluid transitions between clips. 3. **Control Signals**: Text and image prompts are used to guide the generation process, ensuring content, style, and thematic relevance. 4. **Challenges and Solutions**: - **Temporal-Spatial Consistency**: Techniques like adding temporal attention layers and using temporal discriminators enhance temporal-spatial consistency. - **Content Continuity**: Methods such as training directly on long videos and using autoregressive strategies improve content continuity. - **Diversity**: Efforts to generate high-resolution, variable-size, and diverse content are discussed. 5. **Resource Management**: - **Data Compression**: Techniques like T-KLVAE and VAE encoders reduce data dimensionality and computational complexity. - **Lightweight Model Design**: Mask modeling and bidirectional transformers enhance efficiency. - **Training Strategies**: Pre-training on large-scale text-to-image datasets and supplementary dataset methods address resource constraints. 6. **Future Directions**: - **Expansion of Data Resources**: Methods to enrich long video datasets. - **Unified Generation Approaches**: Integrating the strengths of both paradigms. - **Flexible Length and Aspect Ratio**: Enabling generation of videos with variable dimensions. - **Super-Long Videos**: Addressing challenges in generating videos longer than one hour. - **Enhanced Controllability and Real-world Simulation**: Improving understanding and control over generation models. The paper aims to serve as a valuable reference for researchers and practitioners in the field of long video generation, highlighting both current advancements and future research directions.