Script Planner Assistant▸ 
AI Apps 
AI News Alibaba Officially Open-Sources Tongyi Wanxiang Wan2.1 Video Generation AI Model
小编 ·
Alibaba Officially Open-Sources Tongyi Wanxiang Wan2.1 Video Generation AI Model
What is Wan2.1?
Wan2.1 (Tongyi Wanxiang 2.1) is a brand-new open-source video generation model officially released by Alibaba on February 25, 2025. This AI model is open-sourced under the Apache 2.0 license, providing inference code and weights in two parameter specifications, 14B and 1.3B, supporting text-to-video and image-to-video tasks. Its performance has been greatly improved, surpassing other advanced models in multiple benchmark tests, including Sora, HunyuanVideo, Minimax, Luma, Gen3, Pika, and other domestic and international video generation models.
Features of Wan2.1:
Chinese and English Video Generation:
Wan2.1 is the first video model capable of generating text in both Chinese and English, with powerful text generation capabilities, able to produce cinematic-quality text and animations.
It supports font application in various scenarios, including special effect fonts, poster fonts, and fonts displayed in real-world scenes.
Multi-task Support:
Provides text-to-video and image-to-video generation.
Supports tasks such as video editing and video-to-audio.
Text-to-Image (T2I): Supports text-to-image generation.
Video-to-Audio (V2A): Can automatically generate matching audio for videos.
High-Quality Performance:
Built on a hybrid Variational Autoencoder (VAE) and Diffusion Transformer (DiT) architecture, enhancing temporal modeling and scene understanding capabilities.
Capable of simultaneously generating high-definition videos, dynamic subtitles, and multilingual voiceover, supporting 1080p resolution and efficient encoding/decoding.
Consumer-Grade Hardware Support:
The 1.3B version of the Wan2.1 model requires only 8.19GB of VRAM to run, making it compatible with the RTX 4090 graphics card, bringing professional-level video creation into the realm of consumer-grade hardware
Diverse Styles:
Supports a variety of artistic styles, such as cartoon, cinematic color, 3D style, oil painting, classical, and more.
Long Video Generation:
Supports the generation of infinitely long 1080P videos through a feature caching mechanism.
Core Architecture
Hybrid Variational Autoencoder (VAE): Used for video encoding and decoding, supports lossless encoding/decoding of 1080P video streams of any length while preserving temporal information.
Diffusion Transformer (DiT): Based on a video-oriented DiT structure, the Full Attention mechanism ensures effective modeling of long-range spatiotemporal dependencies, achieving spatiotemporally consistent video generation.
Applications of Wan2.1:
Film and Video Production: Quickly create complex scenes and special effects, such as alien worlds in sci-fi films or grand battle sequences, significantly reducing production costs and time.
Advertising and Marketing: Customize creative advertising videos based on brand characteristics to boost brand exposure and appeal.
Education and Training: Produce vivid teaching videos and animated courseware to help students understand knowledge points more intuitively and enrich teaching resources.
Gaming and Entertainment: Generate realistic game scenes and character animations, enhancing game visuals and player immersion.
Short Videos and Social Media: Quickly produce high-quality short videos suitable for social media distribution, effortlessly attracting more viewers.
Virtual Characters and Animation Production: Supports complex character actions and scene design, suitable for virtual character creation in films, advertisements, or games.
Deployment Methods:
Online Experience:
Provides online experience services on Hugging Face Spaces and ModelScope.
Accessible through the Alibaba Tongyi Qianwen APP or website.
Local Deployment:
Detailed local deployment steps are provided, applicable to the 1.3B model.
Ensure that the torch version is 2.4.0 or higher, and install the relevant dependencies.
Use the Hugging Face CLI to download model weights, and deploy via the Gradio framework.
Cloud Image:
Recommended to use the Runninghub platform for online experience of AI applications and workflows.
Newly registered users can receive free credits, supporting cloud image deployment.
Try It Now:
Model Weights:https://huggingface.co/Wan-AI/Wan2.1-I2V-14B-720P
Online Experience:https://modelscope.cn/studios/Wan-AI/Wan-2.1 (recommended version)
ModelScope:https://modelscope.cn/organization/Wan-AI
Github:https://github.com/Wan-Video/Wan2.1
