Wan-Animate-2

Wan-Animate-2

Animates a character picture by copying the motion from a driving video, with text control of the camera viewpoint. Made for people who produce character video and have strong GPUs.

Made by Wan-AI on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Wan-Animate-2 is a character animation model from Wan-AI. You give it a picture of a character and a video with the movement you want, and it makes a new video of that character moving the same way. It skips the usual step of pulling motion data out of the video first.

Who it suits: Animators, video makers and researchers who want to move a character from a still image. Use only pictures and videos you have the right to use, especially when real people appear in them.

Title: Who uses it, and how
Industry | Who uses it | How they use it | Example
Film and video | Animators | Give it character art and a template clip of the motion they want | A character picture performs the movement from a recorded clip
Games | Game artists | Test how character art looks when it moves | A concept image of a character walks and turns in a short clip
Marketing | Brand teams | Animate a mascot picture they own | A mascot image waves in a clip driven by a template video
Research | AI researchers | Study video generation without a separate motion extractor | Compare the Base weights with the Distillation weights

Skip it if you have no strong GPUs, because the default setup assumes a large multi-GPU machine.

Where to use it:

• A Linux machine with several NVIDIA GPUs, set up with conda and Python 3.11
• ComfyUI, Diffusers or DiffSynth-Studio, which the card marks as integrated
• Windows or macOS for downloading the files, which needs Python
• A hosted demo on ModelScope that the card points to

Use case | Who | What they get back
Animate a character | Animators | A video of your character following the template motion
Change the camera angle | Video makers | An output view set by a text prompt, apart from the template video
Streaming animation | Researchers | A Lite variant aimed at real-time character animation

How to set it up: You need a GPU machine, Python 3.11 in a conda environment and Git. The card's defaults are tuned for 8 A800 GPUs at 720P.

• Git, for cloning the code repository with its submodules
• conda with Python 3.11
• NVIDIA GPUs, with 8 A800 for the defaults or 2 A800 for 480P

1. Clone the project's code repository recursively, using the card's clone command.
2. Create the conda environment.
3. Install PyTorch with its vision and audio packages, in the versions the card pins.
4. Install the requirements, then flash-attn, then the repository as an editable package.
5. Download the model weights into a ckpts folder.

conda create -n wan_animate_2 python==3.11 -y

pip install -r requirements.txt
pip install flash-attn --no-build-isolation
pip install -e .

pip install "huggingface_hub[cli]"
huggingface-cli download Wan-AI/Wan2.2-Animate-2-14B --local-dir ./ckpts/

How to use it: First, ask an LLM to caption your reference image with the prompt the card gives, which describes the character's look and the background but not any action. Use that caption as your prompt. Then run the demo script from the infer folder.

export PYTHONPATH="$(pwd)"
cd infer

Tool | What it does
wan_animate_2_demo.py | Runs the generation for the Base model
--prompt | The text description of the character's look and the background
--refer-img-file | The picture of the character to animate
--refer-video-file | The template video whose motion is copied
--config | The YAML file with the parallel settings for your GPUs

Example: You pick a picture of a cartoon kitten and a template video of a person moving. Your prompt describes the kitten's outfit against a plain white background. The script returns a video where the kitten moves like the person in the template.

For newcomers and experts: If you are new, think of it as puppeteering by video. You supply one picture and one clip, and it draws the character moving like the clip. The model needs heavy GPU hardware, so a first small step is to try the ModelScope demo the card links.

For people who work in the field, the model is a redesigned Diffusion Transformer that reads the driving video directly instead of using intermediate motion extractors. The card says this keeps motion detail and the character's identity. Text-driven viewpoint control separates the output camera angle from the driving video. The release has Base and Distillation weights, and a Lite variant targets real-time latency.

Caption the reference picture
Choose a driving video
Run the demo script
Get the animated video

Good to know:

Title: A800 GPUs the card used for each resolution
720P generation: 8 GPUs
480P generation tested: 2 GPUs

• License: Apache 2.0
• Cost to run: no paid service is named, but the card's setup needs several high-end GPUs
• Data it reaches: your image and video files, plus downloads of packages and weights, and a hosted LLM if you choose one for captions
• Not for: machines without strong GPUs, or content you have no right to use
• You may need to change the YAML parallel settings if your hardware differs

On Hugging Face, 13 November 2026

• Downloads in the last 30 days: 0
• Likes: 379
• License: apache-2.0
• Quantized versions: 3
• Fine-tunes: 2

Links

• Model: https://huggingface.co/Wan-AI/Wan2.2-Animate-2-14B
• Files: https://huggingface.co/Wan-AI/Wan2.2-Animate-2-14B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Wan-AI/Wan2.2-Animate-2-14B

Then follow the usage in its model card. Nothing in this kit runs by itself.