How long can you extend a Wan 3.0 video? It depends what you feed it

You have a clip you like and someone has asked for a longer one. The question is whether you can continue what you already have, or whether you are going back to a blank prompt.

You can continue it. Wan 3.0 has a documented extension mode, and it runs forward, backward or both. But how long you can extend a Wan 3.0 video is not a fixed number — the thirty-second limit counts the footage you hand over as well as the footage you get back. Upload twenty seconds and ten is all that is left for you. Upload three and you have twenty-seven.

That single rule is the whole article, and almost nothing written about extension mentions it.

The short answer, and the arithmetic behind it

Thirty seconds is the ceiling for one Wan 3.0 generation, and when a video goes in, input duration plus output duration must total thirty seconds or less. Go over and Wan 3.0 refuses the request outright rather than trimming it.

So the honest answer to "how long can I extend this" is a subtraction:

Every row lands on thirty, which is the point. The ceiling is not something you work up to by extending repeatedly — it is a budget for one call, and your existing footage spends part of it before the model generates anything.

Wan 3.0 has a real extension mode, not a workaround

If you have tried this before, you probably did it by hand: upload the old clip as a reference, describe more of the same, and hope. You do not have to. There is a documented mode with a specific trigger.

You pass the existing video as a reference_video, and the prompt has to carry the intent in words — "extend", "continue", "extend forward", "extend backward". Wan 3.0 reads that as an instruction to continue the footage rather than to be inspired by it. Alibaba recommends leaving the aspect ratio on adaptive so the output keeps the shape of what you sent.

Two capabilities in that mode get almost no coverage:

• Backward extension. You can generate what happens before your clip. If your opening is too abrupt, you do not need a new opening shot — you need three seconds that land on your existing first frame.
• Bidirectional extension. Both ends in one pass.

Neither is a hack. Both are in Alibaba's own description of the mode, which promises that visual style and characters stay consistent while the prompt describes what changes in the new stretch.

Why extending is not the discount it looks like

Here is the part that catches people, and it is worth understanding before you spend anything.

In Wan 3.0, a reference video is billed at your output rate. Reference images are free, audio is free, a document or a link is free — but the seconds of video you upload are charged as though the model generated them. So the meter reads:

billable seconds = uploaded video seconds + generated seconds

Combine that with the thirty-second ceiling and something awkward falls out. Since input plus output cannot exceed thirty, an extension bills at up to thirty seconds no matter how much new footage it returns. What changes is not the price. What changes is how much of it is new.

The last row is the one to sit with. Extending a twenty-five-second clip by five seconds costs the same as generating a fresh thirty-second take. You are paying full price to keep footage you already paid for once.

That is not an argument against extending. Keeping a shot you like is often worth it, and a re-roll is not guaranteed to be as good. It is an argument against assuming extension is the cheap option, which is the assumption most people arrive with.

Feed the tail, not the whole clip

The table above also contains the fix, and it is the single most useful habit here: trim before you upload.

Wan 3.0 needs enough of an ending to continue from — a final state, a camera position, a lighting condition. It does not need the whole take. Cutting your ten-second clip down to its last three seconds before you send it does two things at once: it raises your extension ceiling from twenty seconds to twenty-seven, and it drops the share of the bill spent on footage you already own.

The practical loop looks like this:

1. Export the tail of your clip — the last two to four seconds is a sensible starting range.
2. Send that as the reference video, with adaptive aspect ratio.
3. Write the continuation with the intent word in it, and restate what has to stay the same: the character's wardrobe, the location, the camera axis, the light. Continuation prompts fail the same way multi-shot prompts fail — the model holds what you name and drifts on what you leave out.
4. Join your original and the new stretch in an editor.

Step four is real work and nobody should pretend otherwise. You are matching a cut, not receiving one continuous file.

When you need forty-five or sixty seconds

You cannot get there in one call, and any page telling you otherwise has not read the limit. What you can do is plan the arithmetic instead of discovering it.

A sixty-second piece built from a thirty-second base:

• Round one — a thirty-second generation. 30 billable seconds, 30 seconds of usable footage.
• Round two — upload the last three seconds, generate twenty-seven. 30 billable seconds, 27 new seconds.
• Round three — same shape again. 30 billable, 27 new.

Roughly ninety billable seconds for eighty-four seconds of footage across three files, plus two joins to grade and match. Compare that with the naive plan of extending a full thirty-second clip, which is not possible at all, or extending a twenty-second clip three times, which buys thirty new seconds for ninety billable ones.

Plan the segment lengths before you generate the first one. The cost difference between a well-planned sixty seconds and an improvised one is not marginal, and it is entirely decided before any money is spent.

One more constraint worth knowing early: you cannot cancel a submitted job. A thirty-second generation is committed spend the moment you press the button, which is another reason to settle the plan at 480P first.

The one number you will have to find for yourself

The subtraction and the billing formula above are arithmetic. They come from Alibaba's documentation and from three separate gateways publishing the same rule, so you can check them before you spend anything.

There is one thing you cannot look up: how short a tail can be before continuity breaks. Three seconds is a sensible place to start, not a measured threshold. Identity and lighting may well hold on two seconds for a static shot and need six for a moving one, and a single figure that covers both does not exist.

The good news is that this is the cheapest possible thing to find out for yourself, and you only have to do it once for the kind of footage you make. Take one clip, send its last two seconds, its last three, and its last five as three separate extensions with the same continuation prompt, and watch which one stops holding the face and the light. At 480P that ladder costs a fraction of a single delivery take, and the answer it gives you is about your material rather than about someone else's.

Questions

How long can you extend a Wan 3.0 video?

Up to thirty seconds total, counting the clip you upload. A three-second input leaves twenty-seven seconds of new footage; a twenty-second input leaves ten. There is no setting that lifts the ceiling, and requests that exceed it are rejected rather than trimmed.

Can you extend a Wan 3.0 clip more than once?

Yes, but each round starts the subtraction again, so chaining works only if you upload a short tail each time rather than the whole accumulated piece. Three rounds of "upload three seconds, generate twenty-seven" is a workable pattern; three rounds of uploading everything you have is not.

Does extending cost less than generating a new clip?

Usually not. The seconds of video you upload are billed at the same rate as the seconds Wan 3.0 generates, so an extension charges for the footage you already own. The closer your input is to thirty seconds, the worse that trade becomes — at twenty-five seconds in, you are paying a full generation to add five seconds.

Can Wan 3.0 add footage before an existing clip?

Yes. The extension mode runs backward and bidirectionally as well as forward, so you can generate the seconds leading up to your current first frame. Say so in the prompt with an explicit intent word, and describe the state the new footage has to arrive at.

The short version

How long you can extend a Wan 3.0 video is decided by what you upload, not by a setting: thirty seconds total, minus the clip you send. Extension is not a discount, because uploaded video is billed like generated video. Trim to the tail before you send it, plan multi-segment pieces before you generate the first one, and test the tail length at the cheapest resolution rather than the one you are delivering at.

If you want to see the credit cost of a specific length before you run it, the pricing page computes it from the same numbers the generator does, and the duration guide covers how to choose a length in the first place.