Wan-Move-480P

💡 TLDR: Bring Wan I2V to SOTA fine-grained, point-level motion control!

Wan-Move: Motion-controllable Video Generation via Latent Trajectory Guidance [Paper] Ruihang Chu, Yefei He, Zhekai Chen, Shiwei Zhang, Xiaogang Xu, Bin Xia, Dingdong Wang, Hongwei Yi, Xihui Liu, Hengshuang Zhao, Yu Liu, Yingya Zhang, Yujiu Yang

We present our NeurIPS 2025 paper Wan-Move, a simple and scalable motion-control framework for video generation. Wan-Move offers the following key features:

• 🎯 High-Quality 5s 480p Motion Control: Through scaled training, Wan-Move can generate 5-second, 480p videos with SOTA motion controllability on par with commercial systems such as Kling 1.5 Pro’s Motion Brush, as verified via user studies.

• 🧩 Novel latent Trajectory Guidance: Our core idea is to represent the motion condition by propagating the first frame’s features along the trajectory, which can be seamlessly integrated into off-the-shelf image-to-video models (e.g., Wan-I2V-14B) without any architecture change or extra motion modules.

• 🕹️ Fine-grained Point-level Control: Object motions are represented with dense point trajectories, enabling precise, region-level control over how each element in the scene moves.

• 📊 Dedicated Motion-control Benchmark MoveBench: MoveBench is a carefully curated benchmark with larger-scale samples, diverse content categories, longer video durations, and high-quality trajectory annotations.

🔥 Latest News!!

• Dec 10, 2025: 👋 We've released the inference code, model weights, and MoveBench of Wan-Move.
• Sep 18, 2025: 👋 Wan-Move has been accepted by NeurIPS 2025! 🎉🎉🎉

📑 Todo List

• Wan-Move-480P
  • [x] Multi-GPU inference code of the 14B models
  • [x] Checkpoints of the 14B models
  • [x] Data and evaluation code of MoveBench
  • [ ] Gradio demo

Introduction of Wan-Move

Quickstart

Installation

💡Note: Wan-Move is implemented as a minimal extension on top of the Wan2.1 codebase. If you have tried Wan2.1, you can reuse most of your existing setup with very low migration cost.

Clone the repo:

git clone  https://github.com/ali-vilab/Wan-Move.git
cd Wan-Move

Install dependencies:

# Ensure torch >= 2.4.0
pip install -r requirements.txt

Model Download

Download models using huggingface-cli:

pip install "huggingface_hub[cli]"
huggingface-cli download Ruihang/Wan-Move-14B-480P --local-dir ./Wan-Move-14B-480P

Download models using modelscope-cli:

pip install modelscope
modelscope download churuihang/Wan-Move-14B-480P --local_dir ./Wan-Move-14B-480P

Evaluation on MoveBench

Download MoveBench from Hugging Face

huggingface-cli download Ruihang/MoveBench --local-dir ./MoveBench --repo-type dataset

💡Note:

• MoveBench has provided the video captions. For a fair evaluation, you should turn off the prompt extension function developed in Wan2.1.
• MoveBench provides both data in English and Chinese versions. You can select the language via the --language flag: use en for English and zh for Chinese.

• Single-GPU inference

# For single-object motion test, run:
python generate.py --task wan-move-i2v --size 480*832 --ckpt_dir ./Wan-Move-14B-480P --mode single --language en --save_path results/en --eval_bench

# For multi-object motion test, run:
python generate.py --task wan-move-i2v --size 480*832 --ckpt_dir ./Wan-Move-14B-480P --mode multi --language en --save_path results/en --eval_bench

💡Note:

• If you want to visualize the trajectory motion effect in our video demo, add the --vis_track flag. We also provide a separate visualization script, i.e., scripts/visualize.py, to support different visualization settings, for example, enabling mouse-button effects! 😊😊😊
• If you encounter OOM (Out-of-Memory) issues, you can use the --offload_model True and --t5_cpu options to reduce GPU memory usage.

• Multi-GPU inference

  Following Wan2.1, Wan-Move also supports FSDP and xDiT USP to accelerate inference. When running multi-GPU batch evaluation (e.g., evaluating M