Tutorials

How to Extend a Video with AI and Actually Make It Longer

Looping repeats. Time stretching slows. Continuation generates what happens next, from an mp4 you supply. A practical walkthrough of FLUX 3's continuation lane: what you need, how to chain runs past 20 seconds, why it costs more per second than generating fresh, and the mistakes that waste a run.

2026-09-05 · Kubeez

You shot six seconds. The slot is fifteen. Reshooting is off the table, looping looks like looping, and stretching the clip in an editor just makes it slow.

There is a third option now: hand the footage to a model and let it keep filming. This is a walkthrough of how that actually works, what it costs relative to generating from scratch, and where it falls over.

What "extending a video" actually means

Three different things get called this, and only one of them is what you want.

Looping repeats the clip. Fine for a background texture, obvious the moment anything in frame moves in one direction.

Time stretching slows the footage down and interpolates frames to fill the gap. It fills the runtime, but everything now moves at the wrong speed. Useful for a deliberate slow-motion beat, useless for a longer cut.

Continuation is the one this article is about. The model reads the closing frames of your clip and generates what happens next: same scene, same lighting, same subjects, moving forward in time at normal speed. The result is new footage, not a re-arrangement of the old footage.

FLUX 3 does continuation from a file you supply. That distinction matters. Several video models have an "extend" feature that only continues their own previous generations, referenced by a job id. If the footage did not come out of that model, there is nothing to extend. FLUX 3 takes an mp4 off your disk.

What you need before you start

• An mp4, at most 15 seconds and 50MB. If your source is longer, trim it down to the tail you want continued. The model reads the end of the clip, so the last two seconds are what actually matter.
• A prompt describing what happens next. One line is enough.
• A length, between 5 and 20 seconds. This is required, not optional, and it is what you pay for.

That is the whole list. No timeline, no editing software, no keyframes.

Step by step

1. Open the studio and pick FLUX 3 from the model picker.
2. Attach your mp4. Attaching a video is what puts the request into continuation mode. There is no separate switch to hunt for, and there is no way to be in continuation mode without a video, which is deliberate: the mode and the price are the same decision.
3. Write one line about what happens next. "The camera pushes in past the counter and the barista looks up" beats a paragraph of adjectives. The footage already carries the scene; the prompt only has to carry the change.
4. Pick a length between 5 and 20 seconds. Turn draft mode on first if you are still finding the right prompt. Draft renders at 720p, much faster and much cheaper, and it is for deciding whether the idea works, not for delivery.
5. Run it. Sound is generated with the picture in the same pass, so what comes back has an audio track already, not a silent clip waiting for a foley session.

Making it much longer than 20 seconds

The output of one run is a valid input to the next. That is the part people miss.

Run one gives you 20 seconds. Feed that clip back in and run two gives you 20 more, continuing from where it stopped. Run three, run four. The seams are inside the model rather than in an editor, because each run genuinely reads the frames the previous run produced.

The practical ceiling is not a model limit. It is how many runs you want to pay for, and how far the scene drifts before it stops looking like the same place. In practice, check each run before you commission the next one rather than queueing four in advance.

What it costs, and why continuation costs more

Two things are worth understanding rather than memorising, because rates change and the studio always shows the current one before you run anything.

Billing is per second of output. A 20 second clip costs four times a 5 second clip. There is no flat fee and no minimum beyond the 5 second floor.

Continuing a clip is priced higher per second than generating one from text or a still. Roughly two and a half times higher at the same resolution. That is not a markup decision, it is what the work costs: the model has to ingest and understand your footage before it can produce a frame of its own, and that ingestion is real compute.

Two consequences follow, and both are worth building into how you work:

• Draft mode is not a nicety on this model, it is the correct default while you are still writing the prompt. A draft continuation is a fraction of a full quality one.
• If the shot can plausibly start from a still rather than from footage, starting from the still is meaningfully cheaper. Continuation is for when you specifically need to carry existing footage forward.

The exact per second rate for the setup you have configured is shown in the studio before you commit, and over the API it comes back from the model catalog. Read it live rather than assuming.

Doing it from code

The same continuation lane is available over the Kubeez API and the Kubeez MCP server, with the same limits and the same balance. Upload the mp4 through an upload session, pass it as the source media, set a duration between 5 and 20, and the backend selects the continuation pricing row for you. You cannot accidentally get billed the cheaper text rate for a continuation, and you cannot accidentally pay the continuation rate for a text prompt.

If you are building a pipeline that chains runs, the shape is: generate, poll until complete, take the output URL, feed it in as the source for the next call. Nothing about that requires the browser.

Limits, and the mistakes that cost a run

• Images and a video are mutually exclusive. A request carrying both is rejected before it costs anything, which is the correct behaviour but surprises people who attach a reference still alongside their clip out of habit.
• The input cap is 15 seconds and 50MB. Trim before uploading. Uploading a two minute recording and hoping it takes the tail does not work.
• Draft is 720p only. There is no 1080p draft. Settle the prompt at 720p, then rerun the identical setup at full quality.
• Big changes need to be asked for. Continuation preserves the scene by default. If you want the location to change, the prompt has to say so, and even then a hard cut is usually better done as a separate generation.
• Length is not optional. The model can pick a length on its own, but a per second charge cannot be authorised against a number nobody knows yet, so you set it.

When not to use it

If you need the next shot to be a genuinely different scene, generate it separately and cut. Continuation is for carrying one shot forward, and asking it to teleport is asking it to fight its own strongest property.

If the source clip is already close to what you need and only the ending is wrong, a first and last frame generation may get you there for less. FLUX 3 takes two stills and fills the middle, which is a different tool for a nearby problem.

And if you have no footage at all, you do not need continuation. Start from a prompt or a single still, at the cheaper rate, and continue it later if the finished thing wants to be longer.

The short version

Trim your clip to 15 seconds, attach it, say what happens next, pick a length, run a draft first. Chain runs when you need more than 20 seconds. Expect continuation to cost more per second than generating fresh, and let the studio tell you the number rather than guessing.

Full capability breakdown, including the storyboard and first-and-last-frame modes, is on the FLUX 3 model page.