--- license: apache-2.0 language: - en - zh pipeline_tag: image-to-video library_name: diffusers tags: - video - video-generation --- # Wan-Move: Motion-controllable Video Generation via Latent Trajectory Guidance [](https://arxiv.org/abs/2512.08765) [](https://github.com/ali-vilab/Wan-Move) [](https://huggingface.co/Ruihang/Wan-Move-14B-480P) [](https://modelscope.cn/models/churuihang/Wan-Move-14B-480P) [](https://huggingface.co/datasets/Ruihang/MoveBench) [](https://www.youtube.com/watch?v=_5Cy7Z2NQJQ) [](https://wan-move.github.io/) [](https://www.youtube.com/watch?v=_5Cy7Z2NQJQ) ## 💡 TLDR: Bring Wan I2V to SOTA fine-grained, point-level motion control! **Wan-Move: Motion-controllable Video Generation via Latent Trajectory Guidance [[Paper](https://arxiv.org/abs/2512.08765)]** [Ruihang Chu](https://scholar.google.com/citations?hl=zh-CN&user=62zPPxkAAAAJ), [Yefei He](https://hexy.tech/), [Zhekai Chen](https://scholar.google.com/citations?user=_eZWcIMAAAAJ), [Shiwei Zhang](https://scholar.google.com/citations?user=ZO3OQ-8AAAAJ), [Xiaogang Xu](https://xuxiaogang.com/), [Bin Xia](https://zj-binxia.github.io/), [Dingdong Wang](https://scholar.google.com/citations?user=hRWxWiEAAAAJ), [Hongwei Yi](https://scholar.google.com/citations?user=ocMf7fQAAAAJ), [Xihui Liu](https://xh-liu.github.io/), [Hengshuang Zhao](https://hszhao.github.io/), [Yu Liu](https://scholar.google.com/citations?user=8zksQb4AAAAJ), [Yingya Zhang](https://scholar.google.com/citations?user=16RDSEUAAAAJ), [Yujiu Yang](https://sites.google.com/view/iigroup-thu/about) 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](#quickstart), [model weights](https://huggingface.co/Ruihang/Wan-Move-14B-480P), and [MoveBench](https://huggingface.co/datasets/Ruihang/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 Wan-Move spports diverse motion control applications in image-to-video generation. The generated samples (832×480p, 5s) exhibits high visual fidelity and accurate motion. The framework of Wan-Move. (a) How to inject motion guidance. (b) Training pipeline. The contruction pipeline and statistics of MoveBench. Welcome everyone to use it! Qualitative comparisons between Wan-Move and academic methods and commercial solutions. ## Quickstart #### Installation > 💡Note: Wan-Move is implemented as a minimal extension on top of the [Wan2.1](https://github.com/Wan-Video/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: ```sh git clone https://github.com/ali-vilab/Wan-Move.git cd Wan-Move ``` Install dependencies: ```sh # Ensure torch >= 2.4.0 pip install -r requirements.txt ``` #### Model Download | Models | Download Link | Notes | |--------------|---------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------| | Wan-Move-14B-480P | 🤗 [Huggingface](https://huggingface.co/Ruihang/Wan-Move-14B-480P) 🤖 [ModelScope](https://modelscope.cn/models/churuihang/Wan-Move-14B-480P) | 5s 480P video generation Download models using huggingface-cli: ``` sh pip install "huggingface_hub[cli]" huggingface-cli download Ruihang/Wan-Move-14B-480P --local-dir ./Wan-Move-14B-480P ``` Download models using modelscope-cli: ``` sh pip install modelscope modelscope download churuihang/Wan-Move-14B-480P --local_dir ./Wan-Move-14B-480P ``` #### Evaluation on MoveBench Download MoveBench from Hugging Face ``` sh 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](https://github.com/Wan-Video/Wan2.1?tab=readme-ov-file#2-using-prompt-extension-1) 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 ``` sh # 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](https://github.com/xdit-project/xDiT) USP to accelerate inference. When running multi-GPU batch evaluation (e.g., evaluating MoveBench or a file containing multiple test cases), you should **disable** the [`Ulysses`](https://arxiv.org/abs/2309.14509) strategy by setting `--ulysses_size 1`. Ulysses is only supported when generating a single video with multi-GPU inference. ``` sh # For single-object motion test, run: torchrun --nproc_per_node=8 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 --dit_fsdp --t5_fsdp # For multi-object motion test, run: torchrun --nproc_per_node=8 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 --dit_fsdp --t5_fsdp ``` After all results are generated, you can change the results storage path inside `MoveBench/bench.py`, then run: ``` sh python MoveBench/bench.py ``` #### Run the Default Example For single video generation, (not evaluating MoveBench), we also provide a sample case in the `examples` folder. You can directly run: ```sh python generate.py \ --task wan-move-i2v \ --size 480*832 \ --ckpt_dir ./Wan-Move-14B-480P \ --image examples/example.jpg \ --track examples/example_tracks.npy \ --track_visibility examples/example_visibility.npy \ --prompt "A laptop is placed on a wooden table. The silver laptop is connected to a small grey external hard drive and transfers data through a white USB-C cable. The video is shot with a downward close-up lens." \ --save_file example.mp4 ``` Looking forward to the Gradio launch soon to support everyone in freely creating their own videos. ## Citation If you find our work helpful, please cite us. ``` # Wan-Move: Motion-controllable Video Generation via Latent Trajectory Guidance [](https://arxiv.org/abs/2512.08765) [](https://github.com/ali-vilab/Wan-Move) [](https://huggingface.co/Ruihang/Wan-Move-14B-480P) [](https://modelscope.cn/models/churuihang/Wan-Move-14B-480P) [](https://huggingface.co/datasets/Ruihang/MoveBench) [](https://www.youtube.com/watch?v=_5Cy7Z2NQJQ) [](https://wan-move.github.io/) [](https://www.youtube.com/watch?v=_5Cy7Z2NQJQ) ## 💡 TLDR: Bring Wan I2V to SOTA fine-grained, point-level motion control! **Wan-Move: Motion-controllable Video Generation via Latent Trajectory Guidance [[Paper](https://arxiv.org/abs/2512.08765)]** [Ruihang Chu](https://scholar.google.com/citations?hl=zh-CN&user=62zPPxkAAAAJ), [Yefei He](https://hexy.tech/), [Zhekai Chen](https://scholar.google.com/citations?user=_eZWcIMAAAAJ), [Shiwei Zhang](https://scholar.google.com/citations?user=ZO3OQ-8AAAAJ), [Xiaogang Xu](https://xuxiaogang.com/), [Bin Xia](https://zj-binxia.github.io/), [Dingdong Wang](https://scholar.google.com/citations?user=hRWxWiEAAAAJ), [Hongwei Yi](https://scholar.google.com/citations?user=ocMf7fQAAAAJ), [Xihui Liu](https://xh-liu.github.io/), [Hengshuang Zhao](https://hszhao.github.io/), [Yu Liu](https://scholar.google.com/citations?user=8zksQb4AAAAJ), [Yingya Zhang](https://scholar.google.com/citations?user=16RDSEUAAAAJ), [Yujiu Yang](https://sites.google.com/view/iigroup-thu/about) 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. 🙌 We’re glad to see Wan-Move being tested in real-world videos by many creators and users. ## 🔥 Latest News!! * Dec 15, 2025: 👋 We've released a [local Gradio demo](#gradio-demo) for interactive trajectory drawing and video generation. * Dec 10, 2025: 👋 We've released the [inference code](#quickstart), [model weights](https://huggingface.co/Ruihang/Wan-Move-14B-480P), and...
