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

Problem

Existing motion-controllable video methods, according to the abstract, offer coarse control and scale poorly.

Method

The core idea is to make the condition features themselves motion-aware:

1. represent object motion as dense point trajectories
2. project them into latent space and propagate the first frame’s features along each trajectory
3. use the resulting spatio-temporal feature map as the latent condition for an off-the-shelf image-to-video model (Wan-I2V-14B), without changing its architecture

There’s no auxiliary motion encoder, so fine-tuning scales easily. The authors also build MoveBench, with more data, longer videos and high-quality motion annotations.

Key results

From the abstract: 5-second, 480p videos whose motion control rivals Kling 1.5 Pro’s Motion Brush in user studies, and better motion quality on MoveBench and a public dataset. Code, models and benchmark are released.

Limitations

There’s no dedicated limitations section in the extracted text. The headline comparison with Kling comes from user studies.