About Wan-Move

Welcome to the Wan-Move project. This website provides information about Wan-Move, a motion-controllable video generation framework developed through collaboration between Tongyi Lab at Alibaba Group, Tsinghua University, the University of Hong Kong, and the Chinese University of Hong Kong. The research was accepted at NeurIPS 2025, one of the premier conferences in machine learning and artificial intelligence.

What is Wan-Move?

Wan-Move is a simple and scalable motion-control framework for video generation. It allows users to specify exactly how objects should move in generated videos through dense point trajectories. The system generates high-quality 5-second videos at 480p resolution with motion accuracy comparable to commercial solutions.

The framework introduces latent trajectory guidance, a technique that represents motion conditions by propagating features from the first frame along user-defined trajectories. This approach integrates naturally into existing image-to-video models without requiring architectural changes or specialized motion modules.

Research Team

Wan-Move was developed by a team of researchers from leading institutions:

Key Features

• High-Quality Video Generation: Produces 5-second videos at 832×480p resolution
• Latent Trajectory Guidance: Novel motion control technique using first-frame feature propagation
• Point-Level Control: Dense point trajectories provide precise control over object motion
• Multi-Object Support: Control multiple objects independently with separate trajectories
• No Architecture Changes: Integrates into existing models without specialized modules
• MoveBench Benchmark: Includes a dedicated evaluation benchmark for motion control
• 14B Parameters: Built on the Wan-I2V-14B foundation model

What is MoveBench?

MoveBench is a benchmark dataset introduced alongside Wan-Move for evaluating motion-controllable video generation systems. It includes carefully curated samples with diverse content categories, high-quality trajectory annotations, and visibility masks. The benchmark provides standardized test cases in both English and Chinese, enabling fair comparison of different motion control approaches.

Technical Innovation

The core innovation in Wan-Move is latent trajectory guidance. This technique takes features from the first frame and propagates them along user-defined trajectories. The model learns to generate video content that respects these trajectory constraints while maintaining visual quality and temporal consistency. This approach is simple yet effective, requiring no modifications to the underlying video generation architecture.

Applications

Wan-Move supports various motion control applications:

• Single-Object Motion Control: Guide individual objects along specific paths
• Multi-Object Motion Control: Choreograph multiple objects with independent trajectories
• Camera Control: Simulate camera movements like panning, dollying, and linear displacement
• Motion Transfer: Extract motion patterns from one video and apply to another
• 3D Rotation: Generate videos showing objects rotating in three-dimensional space

How to Use Wan-Move

1. Install Dependencies: Clone the repository and install required packages
2. Download Model Weights: Obtain the Wan-Move-14B-480P checkpoint from Hugging Face or ModelScope
3. Prepare Input: Provide an initial image and trajectory data in NumPy array format
4. Run Inference: Execute the generation script with your inputs
5. Evaluate on MoveBench: Test the model on standardized benchmarks for comparison

Performance

User studies comparing Wan-Move with both academic methods and commercial solutions demonstrate that Wan-Move achieves competitive motion controllability. The system produces videos with accurate motion that follows specified trajectories while maintaining high visual quality. Performance is comparable to commercial solutions like Kling 1.5 Pro's Motion Brush feature.

Open Research

The Wan-Move project follows an open research approach. The code, model weights, and evaluation benchmark are publicly available. This enables researchers and developers to reproduce results, build upon the work, and advance the field of motion-controllable video generation. The comprehensive release includes documentation and examples to help users get started.

Publication Details

• Paper Title: Wan-Move: Motion-controllable Video Generation via Latent Trajectory Guidance
• Conference: NeurIPS 2025
• ArXiv ID: 2512.08765
• Primary Classification: cs.CV (Computer Vision)

Future Development

The research team has indicated plans for future releases, including a Gradio demo interface for easier interaction. Potential future directions include higher resolutions, longer video durations, and additional control mechanisms. The modular design makes the system well-suited for extensions and improvements.