Alibaba Open-Sources Wan2.1-VACE Video Model 69

On the evening of May 14, 2025, Alibaba's Tongyi Wanxiang team open-sourced Wan2.1 - VACE, a unified video editing and generation model. This AI model can not only directly "generate" videos, but also perform various complex editing and re-creation operations on existing videos. Moreover, all of this can be done with a single model, allowing you to precisely control the actions, poses, orientations, composition layout, motion trajectories, styles, and more of characters in the video.

Core Advantages

Wan2.1 - VACE is currently the most comprehensive solution for video generation and editing. A single model architecture can handle numerous video processing tasks, such as text-to-video generation, reference image-to-video generation, video repainting, local video content modification, video background extension, and time extension. It also supports combining multiple tasks, allowing users to complete complex video editing in a single workflow, greatly simplifying the creative process.

In addition, Wan2.1-VACE has another major advantage: users are allowed to freely combine and flexibly mix multiple basic functions to form complex application scenarios.

This model comes in two parameter scales: 1.3B and 14B. The lightweight 1.3B version runs smoothly on consumer-grade GPUs and can output 480P video; the full-power 14B version supports higher resolutions, such as 720P, with better results.

Technical Innovation

This model introduces a brand-new video conditioning unit (VCU), unifying video generation and editing tasks such as text-to-video, reference image-to-video, video-to-video, and local video-to-video. The VCU can transform various inputs, such as text and images, into text, frame sequences, and mask sequences, laying the foundation for the free combination of multiple tasks.

It also processes the frame sequences input to the VCU by dividing them into variable sequences and invariable sequences and encoding them separately, solving the token serialization problem and improving the efficiency of multimodal input processing.

Practical Applications

With this model, you can transform a static vertical image of the "Mona Lisa" into a horizontal dynamic video, and even add glasses to the character, completing three tasks—canvas extension, duration extension, and image referencing—all at once, providing creators with great room for creative expression.

Download Channels

Developers can download and experience this model on platforms such as GitHub, Hugging Face, and ModelScope.

• GitHub: https://github.com/Wan-Video/Wan2.1

• HuggingFace: https://huggingface.co/Wan-AI

• ModelScope: https://www.modelscope.cn/organization/Wan-AI?tab=model

The open-sourcing of Wan2.1 - VACE marks a major step forward in video generation technology, advancing from being capable of only single tasks to combining multiple tasks, providing video creators with more convenient and efficient tools.