Wan-Move: Motion-controllable Video Generation via Latent Trajectory Guidance

Wan-Move represents a new approach to motion-controllable video generation. Developed by researchers from Tongyi Lab at Alibaba Group, Tsinghua University, HKU, and CUHK, this framework introduces a method for generating videos where each element's movement can be precisely controlled through point trajectories. The research was accepted at NeurIPS 2025, demonstrating its significance in the field of video generation.

Traditional video generation models produce content based on text prompts, but controlling the specific motion of objects within those videos has been challenging. Wan-Move addresses this by introducing latent trajectory guidance, a technique that represents motion conditions by propagating features from the first frame along defined trajectories. This allows users to specify exactly how objects should move across the video frames.

The framework builds on the Wan2.1 image-to-video foundation and implements motion control as a minimal extension. This design choice means that existing infrastructure and trained models can be adapted without requiring complete architectural changes or additional motion-specific modules. The result is a practical system that generates 5-second videos at 480p resolution with motion quality comparable to commercial solutions.

What sets Wan-Move apart is its point-level control mechanism. Instead of vague descriptions like "move left" or "rotate," users can define dense point trajectories that specify the exact path each element should follow. This fine-grained control enables applications ranging from single-object animation to complex multi-object choreography, camera movements, motion transfer between videos, and 3D rotations.

The core innovation in Wan-Move is latent trajectory guidance. This technique addresses a fundamental challenge in motion-controlled video generation: how to convey motion information to the model in a way that integrates naturally with existing architectures. The solution is surprisingly simple yet effective.

In Wan-Move, motion is represented by taking features from the first frame of the video and propagating them along user-defined trajectories. Think of it as marking points on objects in the first frame and then telling the model where those points should appear in subsequent frames. The model learns to generate video content that respects these trajectory constraints while maintaining visual quality and coherence.

This approach has several advantages. First, it requires no changes to the underlying video generation architecture. The trajectory information is provided as an additional input condition that guides the generation process. Second, it allows for precise control at the point level, meaning individual parts of objects can be controlled independently. Third, it scales naturally to handle multiple objects, each with their own trajectory paths.

The training pipeline involves pairing video data with trajectory annotations. The model learns to associate the trajectory patterns with the corresponding motion in the video. Once trained, the system can generate new videos where the motion follows user-specified trajectories, producing results that match the intended movement while maintaining high visual quality.

Wan-Move is implemented as a minimal extension on top of the Wan2.1 codebase. This design philosophy offers practical benefits for researchers and developers. If you have previously worked with Wan2.1, most of your existing setup remains usable. The model requires PyTorch 2.4.0 or later and can be installed using standard Python package management tools.

The model weights are available through both Hugging Face and ModelScope platforms. The Wan-Move-14B-480P checkpoint contains the trained parameters for generating 5-second videos at 480p resolution. Download tools are provided through both platforms' command-line interfaces, making it straightforward to obtain the necessary files.

For inference, Wan-Move supports both single-GPU and multi-GPU configurations. Single-GPU inference is suitable for generating individual videos or small batches. For larger-scale evaluation or production use, multi-GPU inference with FSDP (Fully Sharded Data Parallel) provides significant speedup. The system also includes options to reduce memory usage through model offloading and CPU execution of certain components.

Trajectory data is provided in NumPy array format, with separate files for trajectory coordinates and visibility masks. The trajectory file contains the x,y coordinates for each tracked point across all frames. The visibility mask indicates when points are occluded or leave the frame. This straightforward format makes it easy to create custom trajectory data or integrate with motion tracking tools.

The research team has indicated plans for future releases, including a Gradio demo interface that will make the technology more accessible to users without programming experience. This demo will allow users to upload images, define trajectories through an interactive interface, and generate videos with controlled motion.

The current release focuses on 480p resolution and 5-second duration. Future work may explore higher resolutions, longer videos, and additional control mechanisms. The modular design of Wan-Move makes it well-suited for such extensions, as new capabilities can be added without requiring complete redesign of the system.

As the field of video generation continues to advance, motion control will likely become a standard feature in video generation systems. Wan-Move provides a foundation for this capability through its simple yet effective latent trajectory guidance approach. The open release of code and models encourages community involvement in advancing these technologies.