Why extending is not the answer
It would be the obvious move: make one clip, then press "continue" until it is long enough. Two reasons this is not the road to take.
The first is the ceiling. Extension is not unlimited: at one large provider the maximum available is thirty-seven seconds in total, and only at 720p — so a chain does not reach a minute, and you lose resolution on the way there. The second is quality. The vendors write down the side effects themselves: at one of them, text in the prompt that does not belong can cause a cut or a dissolve in the extended part, at another the continuation plays at a different volume from the original and the output comes out a few tenths of a second shorter than asked for.
The edit plan that comes before the generating
This is the heart of the post. AI video does not get longer because you generate more, it gets longer because you work out in advance how many shots it consists of. A minute is not "a minute of video", it is eight to twelve shots, each of them two to eight seconds long.
This is not an AI quirk, it is standard film practice: in a commercial the average shot runs two to four seconds. Which means the eight-second limit is not a limit at all if you think the way an editor thinks. What is missing is not length, it is continuity between the shots — and that can be solved.
The recipe
1. WRITE THE SHOTS DOWN before you generate anything.
One line per shot: what is visible, what moves, how many seconds.
One minute = 8-12 lines. Without this, generating is money burnt.
2. A STARTING FRAME FOR EVERY SHOT.
Same image generator, same style sentence every single time.
This is what decides whether it will look like one video.
3. GENERATE SHOT BY SHOT, IN 2-3 VERSIONS.
Do not accept the first one. In video the spread is wide.
4. BUILD THE SOUND ON A SEPARATE TRACK, one piece end to end.
A continuous music or ambience bed ties the shots together.
It is the cheapest continuity trick in existence.
5. EDIT at the end, in a proper editing program.
That is where the bad shot shows up — and only that one goes again.
The four transitions that work
A change between shots is not distracting when there is a film-grammar reason for it. Four solutions that sit well with AI clips too.
- Cutting on action. Something starts moving at the end of one shot and carries on at the start of the next. The eye follows the movement and does not register the cut.
- A jump in shot size. Wide to close. A cut between two similar sizes jumps; a cut between two very different ones does not.
- A wipe-by. Something crosses the frame and covers it — an arm, a car, a door. The next shot starts from there. This is the most reliable one, because it hides the differences.
- A sound bridge. The audio of the next shot comes in half a second before its picture. Simple, and every editing program has it.
Four grips on continuity
- Make the last frame of one shot the starting frame of the next. This is the strongest method, and most providers support it. The two clips then physically start from the same place.
- A closing frame as well, where possible. Where the end can be specified too, the motion runs between two fixed points instead of into nothing.
- The same seed. It helps, but do not trust it: one vendor writes word for word that it "does not guarantee determinism, it only slightly improves it."
- The same reference set. If there is a character in the scene, this is the decisive one — and character consistency on video walks through the method step by step.
What this costs in time and money
Let us count a one-minute video with twelve shots, in the mid price bracket. Generating takes two or three attempts per shot, so twenty-five to thirty clips. At ten to twenty cents a second that comes to twenty to forty dollars of raw generation cost.
The time, though, does not come from the generating: that part is minutes. The work is the planning (half an hour), the starting frames (an hour), the selection and the edit (one to two hours). One minute of finished AI video is realistically half a day — still a fraction of a shoot, but not the "type it and it is done" experience the demos suggest.
Common misconceptions
- "I will solve it with the extend feature." It has a ceiling, you lose resolution, and the side effects the vendors document come along with it.
- "A good prompt is enough and the model will do the whole sequence." Getting several shots out of one request exists at one or two providers today, and even there in a limited form.
- "The edit will sort itself out at the end." The other way round: generating without an edit plan is the most expensive mistake in this workflow.
- "If I use the same prompt, I get the same thing." You do not. Continuity comes from the starting frame and the reference, not from the prompt.
What to take away
A one-minute AI video is an editing job, not a generating job. Plan the shots in advance, make a starting frame for every one of them, and build the sound as a single piece — that is what turns it into a video instead of a pile of clips. Why the picture-first route is more predictable is the subject of text-to-video or image-to-video, and the camera moves inside each shot are handled by camera language in the prompt, the next post along.