1. Video Input

Load Video
In this step, you load your source video into the workflow. The frame_load_cap is set to the recommended 81 frames, following WanVideo's guidelines. However, you’re free to adjust this value and experiment with different settings to suit your needs.

Resize Image
This node allows you to adjust the resolution of your video. Choosing a lower resolution will reduce the computational load and speed up the entire process — especially helpful if you're using less powerful hardware.

Start Frame

Here you insert the start frame you created in the inpainting workflow. Make sure this frame exactly matches the first frame of your video. This ensures that the new character stays aligned with the original video for a consistent final result.

Prompt

Enter your prompt here to define the appearance and actions of the new character.
We recommend describing both the look and the intended movement or pose in detail.

A clear and precise prompt ensures that the new character not only matches visually but also fits naturally into the movement and action of the video.

ControlNets

In this section, you’ll find the Fast Group Bypasser. This tool lets you selectively enable specific ControlNet groups, clearly labeled like "CN01_POSE" or "CN02_01_FACEMASH." Activate only the groups you need to achieve your desired effects. We'll go into more detail about the different groups in the next section.

You’ll also see the Video Combine setting. The frame rate is intentionally set to 16 FPS, because WanVideo internally operates at 16 FPS. Please keep this setting unchanged to avoid potential issues during sampling.

CN01_POSE

This group uses the DWPoseEstimator to convert body movements from the video into a pose schema. Pay attention to the enabled detection modes, such as detect_hand, detect_body, or detect_face.

If your character doesn’t have a humanoid face, "Detect Face" might not work well. In our film Samsa, we disabled this function because it was incompatible with the character design.
Adjust the settings according to your specific needs.

CN02_01_FACEMESH

This section is disabled by default. It becomes useful if you want to try a different method for lipsync. As you can see, only the mouth feature is currently enabled for lipsync.
However, feel free to experiment with other options to see what works best for your needs.

CN03_01_LINEART

In this step, the environment around the tracked pose is converted into Lineart, while the area of the replaced character is masked out.
This allows us to extract camera movements and background motion, and later apply them back to the start frame for consistent camera behavior.

Be sure to adjust the Grow Mask with Blur settings (size and blur radius) to fit the new character properly.

We deliberately avoid including Lineart over the replaced person to prevent transferring proportions or details from the original character to the new one.

CN04_DEPTH

At this stage, a Depth Map captures the depth information of the background and applies it to the start frame. This helps maintain realistic depth and movement in the generated video.

However, be cautious:
If your new character differs significantly from the original — for example, with a very different hairstyle — the Depth Map might introduce unwanted distortions.

05_WanVideo

In this section, you’ll find the KSampler.
The step count is currently set to 25, but you can adjust it as needed:

• More steps (e.g., 30) → higher accuracy

• Fewer steps (e.g., 20) → faster results, sometimes even more creative outputs

Higher steps make ControlNet applications more precise, while fewer steps might introduce creative variations. Adjust these settings to find the perfect balance between accuracy and creativity for your project.