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

Wan-Move is a framework for motion-controllable video generation via latent trajectory guidance, providing advanced, point‑level motion control for image‑to‑video synthesis. It achieves high‑quality 5‑second 480p video generation with industry‑leading motion control capabilities, even comparable to commercial systems.

Wan‑Move introduces a novel latent trajectory guidance mechanism, which represents motion conditions by propagating first‑frame features along trajectories. This approach integrates seamlessly into existing image‑to‑video models without architectural changes or additional motion modules. It supports fine‑grained point‑level control, allowing precise motion manipulation of each element in the scene through dense point trajectories. To advance the field, Wan‑Move releases a dedicated motion‑control benchmark dataset called MoveBench, which contains large‑scale samples, diverse content, longer video durations, and high‑quality trajectory annotations, enabling comprehensive evaluation of motion‑control performance.

Key Features of Wan‑Move

• High‑Quality Motion Control: Trained at scale to generate 5‑second 480p videos with industry‑leading motion control. User studies validate that its performance matches commercial systems such as Kling 1.5 Pro’s Motion Brush.
• Innovative Latent Trajectory Guidance: The core idea is to propagate features from the first frame along trajectories to represent motion conditions. This integrates seamlessly into off‑the‑shelf image‑to‑video models (e.g., Wan‑I2V‑14B).
• Fine‑Grained Point‑Level Control: Uses dense point trajectories to describe object motion, enabling precise region‑level control over how each element in the scene moves.
• Dedicated Motion‑Control Benchmark – MoveBench: Includes larger‑scale samples, longer video durations, high‑quality trajectory annotations, and covers a wide range of content categories, making it suitable for evaluating motion‑control models.

Application Scenarios of Wan‑Move

Wan‑Move supports various motion‑control applications, generating videos (832×480p, 5 seconds) with high visual fidelity and accurate motion effects, including:

• Multi‑object motion control
• Complex motion simulation
• Combined object and camera motion
• Basic‑level motion control
• Motion transfer
• 3D rotation effects

For more information about Wan‑Move, visit wan-move.github.io.