MiniMax H3 Max Video Insert: How to Use It, Timing Rules and Cost

You have a clip you like. Then you notice the shot that's missing: the product close-up nobody filmed, the reaction that would sell the moment, the beat where something should actually happen. Until now the fix was a reshoot, or generating a separate clip and hoping it cuts in cleanly.

MiniMax H3 Max Video Insert does it in one step. You pick a moment in your video, describe what should happen, and it generates a new 5 to 13 second scene that starts there and then hands back to your original footage. You get one finished video, not a loose clip to edit in.

It went live on September 28. We added it to Veevid as the AI Video Inserter the next day, and before shipping it we ran 27 test requests, including a batch designed to fail, to find out where the real limits are. Most of what's below comes from those runs rather than the launch notes.

What Video Insert Actually Does

There are two ways to use it, and they only differ in what happens to your footage around the new scene.

• Insert: Your video plays up to the cut point, the new scene plays, then the rest of your video continues. Your footage is kept (apart from a single frame, explained below), so the finished video is your original length plus the new scene.
• Replace: You set a cut point and a resume point. Everything between them is swapped for the new scene, and your video picks up from the resume point.

Either way the output is a single continuous file. Your footage outside the new scene isn't regenerated, and color match (on by default) grades the new scene to sit in the same light as the rest.

Here's an insert. The original is a drone shot of a car on a snowy track. The prompt asked for a polar bear to lumber onto the track and the car to swerve around it:

And a replace. The original knight clip is a slow walk toward a ruined archway. We replaced part of the walk with a spectral wolf bursting out of the arch:

How to Use Video Insert on Veevid

On Veevid the whole flow sits on one page:

1. Upload your video. MP4, MOV, WebM or M4V, about 3 to 60 seconds long and up to 50MB.
2. Mark where the new scene goes. Drag the marker on the timeline and the preview shows the exact frame. Leave "Keep all original footage" on to insert, or turn it off and set a resume point to replace a section.
3. Describe the new scene. Write only what happens in the new shot. You can add up to 9 reference images and 3 reference videos, and the prompt becomes optional once you do.
4. Pick a length and resolution, then generate. The new scene can be 5 to 13 seconds, at 480p or 768p. The full price is on the Generate button before you spend anything.

In our runs a 5 second scene came back in roughly two minutes, including upload and saving to your account.

Try it on the AI Video Inserter

The Rules We Found by Testing

The model has a few timing constraints that aren't in the launch notes, and getting one wrong means a failed request. Here's what our test batch confirmed:

A few of these are worth explaining.

The end margin isn't a fixed number of seconds from zero. It's measured from the end of your clip. On a 10.04 second video the model accepted a resume point up to 8.667 seconds. Browsers sometimes report a clip as slightly longer than its picture track, so the timeline on Veevid keeps an extra quarter second of margin and only lets you drag to points that will actually work. You won't burn a request on a setting the model will reject.

"Insert" is really a one-frame replace. The model always works between a cut point and a resume point. To add a scene while removing only one frame, the resume point sits one frame (1/24 of a second) after the cut. The toggle handles this for you, and it's why insert mode still shows a resume point one frame after the cut.

Resolution follows your choice, not your source. A 720p source rendered at 768p came back as 1344x768 for the whole video, so the original footage is scaled along with the new scene.

What to Expect From the Results

The new action doesn't always fill the whole scene. The model eases in and out of your footage, so the main event usually lands in the middle of the new scene. In our perfume example below, the close-up itself runs about three seconds of a five second scene, with the rest spent cutting in and back out. If you need a longer beat, ask for a longer scene rather than a busier prompt.

It isn't a silent patch. The finished file keeps your original audio, and the new scene comes with sound of its own. We measured loudness on a production run to check: the inserted section sat at the same level as the rest of the clip rather than dropping out.

Reference images are how you control specifics. The perfume bottle above came from a single product photo. Without a reference, the model invents its own version of whatever you describe, which is fine for a polar bear and useless for your product.

Two settings change the feel. Prompt enhancement rewrites your prompt using your video so the scene fits in; turn it off when you need the prompt followed exactly. Color match keeps the new scene in your footage's grade; turn it off when you want the scene to look deliberately different, like a flashback.

How the Cost Works

You pay for the new scene, not your original video. Its length doesn't change the price, so inserting 5 seconds into a 58 second clip costs the same as inserting it into a 10 second one. We checked this directly with a near-60 second source.

The main things that move the price are scene length and resolution, with 768p costing more per second than 480p. Up to four reference images with no reference video cost nothing extra. Reference videos are billed by length and add up quickly: at 768p even a few seconds of reference video can cost more than the scene itself. Veevid includes all of it in the price shown on the Generate button, so check it before you generate.

That makes 480p the sensible way to test an idea, then rerun at 768p once the timing and prompt are right.

Where It's Useful

• Product shots you didn't film: Add a close-up of the product in someone's hands to an existing lifestyle clip, using a product photo as the reference.
• Missing pickups: Cover the reaction, the glance or the gesture that would have made a scene work.
• B-roll inside the shot: Break up a long take with a cutaway without leaving the clip.
• Plot beats and effects: Drop in something that happens, like the bear or the wolf above, while the setting stays yours.
• Ad variants: Keep the same base footage and insert a different moment for each version.

Prompt Tips

• Describe only the new scene. Don't retell what happens in your footage; the model already sees it.
• Name the transition when it matters. "Cut to a close-up of her hands" gets a different shot from "she lifts the bottle".
• Keep one main action. Five seconds holds one clear event well. If you need two, make the scene longer.
• Match the camera if you want it seamless. Mentioning that the drone keeps following the car helps the new scene flow into the original motion.

Get Started

Video Insert is live on Veevid as the AI Video Inserter. If you want the rest of the model, MiniMax H3 Max also covers text to video, image to video, camera controls, lip sync and extending a clip with the AI Video Extender. For how H3 Max compares with the standard model, see MiniMax H3 vs H3 Max.