What is Wan-Move
Wan-Move is a motion-controllablevideo generationframework open-sourced by institutions such as Alibaba Tongyi Lab, achieving high-quality video motion control through latent trajectory guidance. The core highlight is that it requires no architectural changes to existing image-to-video models, enables fine-grained point-level motion control, can generate 5-second, 480p videos, and its motion control quality is comparable to commercial systems. Wan-Move provides the MoveBench benchmark, which includes large-scale samples and high-quality trajectory annotations, and can be used to evaluate and compare the motion control capabilities of different methods.
Main Features of Wan-Move
-
High-quality motion control: It can generate 5-second, 480p videos, and its motion control quality is comparable to commercial systems, meeting the needs of high-quality video creation.
-
Latent trajectory guidance: By propagating the features of the first frame along trajectories to generate aligned spatiotemporal feature maps, no additional motion module is needed, and it can be seamlessly integrated into existing image-to-video models.
-
Fine-grained point-level control: It supports precise region-level motion control for each element in the scene, enabling highly customized video effects.
-
MoveBench benchmark: It provides large-scale, diverse, long-duration video samples and high-quality trajectory annotations for evaluating and comparing the motion control capabilities of different methods.
-
Open source and ease of use: The code, model weights, and MoveBench are all open-sourced, allowing users to quickly get started with video generation and motion control experiments, lowering the barrier to use.
Technical Principles of Wan-Move
-
Latent trajectory guidance: Wan-Move uses latent trajectory guidance technology to represent object motion as dense point trajectories and map these trajectories into latent space. By propagating the features of the first frame along the trajectories, it generates aligned spatiotemporal feature maps, thereby achieving fine control over motion in video generation.
-
No additional modules needed: The framework can achieve motion control without architectural changes to existing image-to-video models (such as Wan-I2V-14B) and without adding additional motion encoders, offering excellent compatibility and extensibility.
-
Large-scale training: Through large-scale training data and optimization, Wan-Move can generate high-quality 5-second, 480p videos, and its motion control capability is comparable to commercial systems, ensuring the visual quality of videos and the accuracy of motion.
-
Fine-grained point-level control: Object motion is represented by dense point trajectories, allowing users to perform precise region-level motion control for each element in the scene and achieve highly customized video effects.
-
MoveBench benchmark: To validate and evaluate motion control performance, Wan-Move provides the MoveBench benchmark, which includes large-scale samples, diverse content categories, long-duration videos, and high-quality trajectory annotations, providing a standardized testing platform for research and development.
Project Address of Wan-Move
- Project official website：https://wan-move.github.io/
- Github repository：https://github.com/ali-vilab/Wan-Move
- HuggingFace model library：https://huggingface.co/Ruihang/Wan-Move-14B-480P
- arXiv technical paper：https://arxiv.org/pdf/2512.08765
Application Scenarios of Wan-Move
-
Video creation: Users can generate videos with specific motion effects by defining the motion trajectories of objects, making it suitable for fields such as animation production, special effects design, and creative short videos, helping creators quickly realize complex motion scenes.
-
Advertising and Marketing: In advertising videos, Wan-Move can be used to generate dynamic product displays, brand stories, and more, attracting viewers' attention through fine-grained motion control and enhancing the appeal and influence of advertisements.
-
Video Editing: It supports editing the first frame of a video and applying these changes to the entire video. It can also perform motion replication and camera motion control, helping video editors quickly adjust and optimize video content.
-
Education and Training: In the field of education, Wan-Move can generate instructional videos with dynamic demonstration effects, such as physics experiments and biology animations, helping students better understand and learn knowledge through vivid motion displays.
-
Game Development: In game development, Wan-Move can be used to generate animation effects in games, such as character actions and scene changes, enhancing the game's visual effects and user experience.
-
Virtual Reality (VR) and Augmented Reality (AR): Wan-Move can generate dynamic video content that matches virtual environments or augmented reality scenes, providing users with a more immersive experience.