Turn an image into a video with Wan 2.2 local model

Wan 2.2 is a local video model that can turn text or images into videos. In this article, I will focus on the popular image-to-video function. The new Wan 2.2 models are surprisingly capable in motion and camera movements. (But not necessarily controlling them)

This article will cover how to install and run the following image-to-video models:

1. Wan 2.2 14B model – higher quality but needs more VRAM and takes longer.
2. Wan 2.2 5B model – Slightly lower quality but is quicker and needs lower VRAM.

Software needed

Wan 2.2 model

Wan 2.2 is a multimodal AI model from Wan AI, a highly ancipated update to Wan 2.1. It uses a new “Mixture of Experts” model architecture, where different expert sub-models handle different levels of noise during video creation. The result is cleaner, sharper videos.

Main features:

• Movie‑quality visuals: You can adjust lighting, color, and framing using a prompt.
• Smooth complex motion: Good improvement in motion generation.
• Accurate scene understanding: Generating scenes with many objects and details.

You can create videos from text or images.

The models and the code are open-source and under the permissive Apache 2.0 license, making it open to commercial use.

image-to-video with the Wan 2.2 14B model

This workflow turns an input image into a video. It requires 20 GB VRAM and takes about 1 hr 20 mins on an RTX4090 GPU card.

Step 0: Update ComfyUI

Before loading the workflow, make sure your ComfyUI is up to date. The easiest way to do this is to use ComfyUI Manager.

Click the Manager button on the top toolbar.

Select Update ComfyUI.

Restart ComfyUI.

Step 1: Load the workflow

Download the workflow below.

Step 2: Download models

After loading the workflow JSON file, ComfyUI should prompt you to download the missing model files.

Here are what you need to download:

• Download wan2.2_i2v_high_noise_14B_fp8_scaled.safetensors and put it in ComfyUI > models > diffusion_models.
• Download wan2.2_i2v_low_noise_14B_fp8_scaled.safetensors and put it in ComfyUI > models > diffusion_models.
• Download umt5_xxl_fp8_e4m3fn_scaled.safetensors and put it in ComfyUI > models > text_encoders.
• Download wan_2.1_vae.safetensors and put it in ComfyUI > models > vae.

Step 3: Upload an image for the start frame

Upload an image you wish to use as the first frame of the video.

You can use the test image below.

Step 4: Revise the prompt

Describe your video in the prompt. Use keywords to direct the camera.

Step 5: Generate the video

Click the Run button to run the workflow.

image-to-video with the Wan 2.2 5B model

This workflow turns an input image into a video. Unlike the 14B model, this smaller model requires only 8 GB VRAM and takes about 6 mins on an RTX4090 GPU card.

Step 1: Download the workflow

Download the workflow below and drop it into ComfyUI to load.

Step 2: Install models

Here are the model files you need to install:

• Download wan2.2_ti2v_5B_fp16.safetensors and put it in ComfyUI > models > diffusion_models.
• Download umt5_xxl_fp8_e4m3fn_scaled.safetensors and put it in ComfyUI > models > text_encoders.
• Download wan2.2_vae.safetensors and put it in ComfyUI > models > vae.

Step 3: Upload an image for the start frame

Upload an image you wish to use as the first frame of the video.

You can use the test image below.

Step 4: Revise the prompt

Describe your video in the prompt. Use keywords to direct the camera.

Step 5: Generate the video

Click the Run button to run the workflow.

Testing camera keywords on Wan 2.2

The Wan AI team touted a significant improvement in the model’s ability to follow camera instructions. Let’s put it to the test. I will stick with the 5B model due to the generation time.

Zoom

Zooming out from a character is working pretty well:

Sometimes, it confuses zooming in and out.

The solution is to generate a few more videos and pick the best one.

But it is not clear whether the zoom is working when the character is doing something at the same time.

Pan

Panning right works pretty well:

Panning left is still producing a video of panning right (and zooming out).

I think what happens is that the initial image is more natural for the video to pan right than left. The model produces likely videos rather than following the prompt.

Useful links

Wan2.2 Day-0 Support in ComfyUI – Release press from ComfyUI

Wan-Video/Wan2.2 – Github page hosting the source code

Wan-AI/Wan2.2-I2V-A14B · Hugging Face – Model weights of the 14B I2V model

Wan-AI/Wan2.2-TI2V-5B · Hugging Face – Model weights of the 5B I2V model

Wan2.2 Video Generation – ComfyUI’s official documentation