Wan2.2-I2V-A14B

Wan2.2-I2V-A14B

Turns a still image into a video at 480P or 720P, using a Mixture-of-Experts design from Wan-AI. For creators and researchers who generate video from pictures.

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: Wan2.2-I2V-A14B is the image-to-video model in the Wan2.2 family from Wan-AI. It supports 480P and 720P output. Its Mixture-of-Experts design aims for steadier video, fewer unrealistic camera moves and better handling of stylized scenes.

Who it suits: Video creators, researchers and developers who want to animate a still picture with an open model they run themselves. Prompts in English and Chinese are listed as supported.

Industry | Who uses it | How they use it | Example
Video production | Video creators | Animate a still frame into a short clip | Turn a concept image into a moving shot
Research | Video generation researchers | Study a Mixture-of-Experts video diffusion model | Run the model and compare it with Wan2.1
Tooling | Developers | Build image-to-video into their own apps or workflows | Call the model from the project's inference code
Stylized content | Illustrators | Bring stylized scenes to life | Animate an illustrated scene at 720P

You can skip it if you only have a text prompt and no starting image. The same family has separate text-to-video models.

Where to use it:

• Windows, macOS and Linux with Python, through the project's own code
• ComfyUI, which the card says integrates Wan2.2
• Diffusers, through the separate Diffusers version of this model
• Multi-GPU setups, since the card ships multi-GPU inference code

Use case | Who | What they get back
Animating a picture | Creators | A video clip that starts from your image
Higher-resolution output | Producers | Video at 480P or 720P
Research comparison | Researchers | Output to compare against Wan2.1

How to set it up: You need Python, the project's code and its dependencies.

• Python, with torch version 2.4.0 or newer
• The Wan2.2 code repository from GitHub
• The model weights, downloaded from Hugging Face

1. Clone the Wan2.2 repository from GitHub and move into its folder.
2. Install the dependencies. If flash_attn fails to install, install the other packages first.
3. Download the model weights.

cd Wan2.2

# Ensure torch >= 2.4.0
# If the installation of `flash_attn` fails, try installing the other packages first and install `flash_attn` last
pip install -r requirements.txt

pip install -U huggingface_hub
hf download Wan-AI/Wan2.2-I2V-A14B

How to use it: The card's run steps go on to a model download and a generation command, which are cut off in the part shown here. Take the exact command from the full card. You provide a starting image and a text prompt, and choose 480P or 720P.

Example: You give it a picture of a mountain village and a prompt about clouds drifting and smoke rising. It returns a short video that starts from your picture and adds that motion at the resolution you chose.

For newcomers and experts: Image-to-video means you start with one picture and the model works out the frames that follow. A first small step is to get the repository running with its dependencies, then try one image at 480P.

The design splits the denoising process across timesteps between specialized expert models. The card says this adds capacity without raising the computational cost. Wan2.2 also trained on more data than Wan2.1, and added aesthetic labels for lighting, composition, contrast and color tone. The 5B hybrid model with the 16x16x4 compression VAE is a separate checkpoint, not this one.

Start from one still image
Add a text prompt for the motion
Expert models denoise in turn
Get a 480P or 720P video

Title: Extra training data over Wan2.1
Images: 65.6%
Videos: 83.2%

Good to know:

• This repository holds the 14B-class image-to-video checkpoint. The 5B model is a separate release.
• The card lists ComfyUI and Diffusers integrations as done.
• Installing flash_attn can fail, so the card suggests installing it last.
• License: apache-2.0.
• Cost to run: none named, but multi-GPU inference code hints at heavy hardware.
• Data it reaches: files from Hugging Face and GitHub, plus the images and prompts you supply.
• Not for: the card names no limits.

On Hugging Face, 13 November 2026

• Task: image to video
• Library: wan2.2
• Formats: Safetensors
• Downloads in the last 30 days: 8,500
• Likes: 959
• License: apache-2.0
• Languages: en, zh
• Quantized versions: 59
• Fine-tunes: 143
• Tags: wan2.2, diffusers, safetensors, image-to-video

Links

• Model: https://huggingface.co/Wan-AI/Wan2.2-I2V-A14B
• Files: https://huggingface.co/Wan-AI/Wan2.2-I2V-A14B/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-I2V-A14B

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