Video Inpaint

Video Inpaint lets you edit an existing video with words. Swap a subject, change an outfit, replace an object, restyle part of the scene — while everything else in the footage stays exactly as it was. There's no frame-by-frame masking and no rotoscoping: you simply name what you want changed, and MiniMax H3 finds it, tracks it through the whole clip, and repaints it.

How It Works

1. In the MiniMax H3 tab, select the Inpaint task.

2. Upload your source video — the clip you want to edit (up to 15 seconds; longer uploads can be trimmed in the picker).

3. Type the Video Mask Prompt — a short phrase naming what should be replaced (e.g. "the red car").

4. Pick a control type (see below).

5. Optionally add up to 5 reference images to show exactly what the replacement should look like.

6. Type the Prompt — a description of what should appear instead. Recommend to use AI Writer to help you write once and change the prompt based on the structure.

The result keeps the source clip's length and aspect ratio automatically — that's why the duration slider and aspect ratio control disappear for this task.

What Is the Video Mask Prompt?

In classic image inpainting, you paint a mask over the area you want to change. In video, painting a mask on every frame would be impossibly tedious — so H3 masks from text instead. The Video Mask Prompt is that text: it tells the model what to find in the footage, and H3 detects it and follows it across every frame, even as it moves, turns, or is partially hidden.

The two text fields do different jobs, and keeping them straight is the key to this task:

Field	Answers	Example
Video Mask Prompt	What in the clip gets replaced?	the woman's jacket
Prompt	What appears in its place?	a black leather biker jacket with silver zippers

Think of the mask prompt as pointing, and the prompt as describing.

More Example:

Use Case	Video Mask Prompt
Replace the entire person	person
Change the face and hair of a person	head, hair
Replace the woman on the right	woman on the right
Replace a red car	red car

Control Type

• None — the masked region is freely regenerated from your prompt. Use this when replacing an object with something that moves differently (a car → a horse, a coffee cup → a crystal ball).

• Pose — the original motion is preserved as much as possible while the appearance changes. Use this when the performance matters: swapping a person's identity or outfit while keeping their exact dance, or gestures.