Script Planner Assistant▸ 
AI Apps 
AI News Wan 2.1: Open-Source Video Generation Model by Tongyi Wanxiang
小编 ·
Wan 2.1: Open-Source Video Generation Model by Tongyi Wanxiang
The Tongyi Wanxiang team has released Wan 2.1 (WanXiang 2.1), their latest video generation model. This model has made significant advancements in multiple aspects of video generation, including video quality, complex motion representation, physics simulation, and Chinese character generation. Particularly in Chinese content generation, it is the first video model capable of effectively controlling the generation of Chinese text effects, addressing the shortcomings of previous models in handling Chinese content.
Wan 2.1 Features
Complex Motion Representation: Wan 2.1 can stably render various complex human body movements such as spinning, jumping, and tumbling, making the videos look more vivid and realistic.
Physics Simulation: It realistically replicates the physical laws of the real world, making the generated videos more authentic and aligned with user expectations.
High-Quality Video Generation: The videos generated by this model feature more realistic visual effects and improved pixel quality. It enables the generated content to follow physical principles and executes complex instructions with greater precision.
Complex Motion Handling: Wan 2.1 excels at handling complex motion scenes, such as figure skating, swimming, and diving. It keeps movements coordinated and realistic, achieving significant improvements in generating complex body motions.
Multilingual Support: Wan 2.1 is the first video generation model that simultaneously supports both Chinese and English text effects, greatly expanding its application scope in the creative industry—including advertising and short video production—meeting both domestic and international needs.
Efficient Encoding and Decoding Technology: It employs its self-developed VAE (Variational Autoencoder) and DiT (a deep learning architecture) technologies, making it possible to generate infinitely long 1080P videos while improving both efficiency and image quality.
Wan 2.1 Technical Architecture
Wan 2.1 adopts its self-developed VAE (Variational Autoencoder) and DiT (Diffusion Transformer) architecture, which significantly enhances its ability to model spatiotemporal contextual relationships. The VAE splits videos into multiple blocks and caches intermediate features, enabling efficient encoding and decoding of ultra-long 1080P videos and supporting the generation of infinitely long high-definition videos. The DiT leverages spatiotemporal full-attention mechanisms and parameter-sharing mechanisms, significantly improving generation quality and training efficiency.
The Wan 2.1 model also supports infinitely long video generation. It uses VAE to split videos into multiple parts and cache intermediate features for efficient encoding and decoding, thus supporting video generation at 1080P length. In addition, its generation quality and training efficiency are significantly enhanced by the spatiotemporal full-attention mechanism and parameter-sharing mechanism, resulting in more refined and stable outputs. Furthermore, when setting parameters, users can choose different preset modes, such as simple mode and professional mode, to meet different creative skill levels and project requirements.
Wan 2.1 Application Scenarios
Content Creation: For video creators, Wan 2.1 is a powerful tool that helps them create videos with richer effects.
Advertising Production: In marketing and advertising, it can be used to create highly engaging promotional videos.
Education and Training: In the education field, it enhances the learning experience through vivid demonstration videos and improves the efficiency of information delivery.
How to Use Wan 2.1
Access the Platform: Visit http://tongyi.aliyun.com/wanxiang/ to register an account and log in.
Select a Feature: After logging in, find "Text-to-Image" (文字作画) in the left sidebar and select the Wan 2.1 feature, where you can see different generation options.
Enter Prompts: Users can enter the topic or elements they want to generate in the video. Both Chinese and English inputs are supported.
Set Parameters: Users can adjust video generation parameters according to their needs, such as video length, effect types, and character movements.
Generate and Preview: After submission, the system will generate the video. They can review the preview and download or share it once satisfied.
Demo URL:https://tongyi.aliyun.com/wanxiang/
