Generating video with AI

What one generation actually hands you

One short video file, uncut, with no memory of the generation before it. That is the whole delivery.

A video model builds frames from your sentence and, like an image model, finds nothing. What separates it from an image model is duration: the output covers a span of time, and that span is short today. Seedance 2.5 gives up to 30 seconds in one generation, Kling VIDEO 3.0 up to 15, Runway Gen-4.5 up to 10 and Veo 3.1 exactly 8. Google own documentation for Veo is even more specific: one video per request, 24 frames per second, and the 8 second length is only allowed at 1080p or 4k or when you use reference images.

The practical consequence: your one minute video is not one generation. It is four to eight separate generations that then have to be assembled. And because these models hold no memory between generations, every press of the button can change the colour of a jacket, the quality of the light or a person face. The Veo documentation confirms this in a blunt sentence: referencing or reasoning across multiple videos is not supported, and attempting it degrades the output.

So the real skill here is not prompt writing. It is cutting an idea into shots that each fit under the model ceiling and still match each other. Nearly every tutorial we have seen concentrates on the first half and drops the second, and the second half is exactly where professional work separates from a pile of clips.

One more constraint that never makes a headline: sound. Veo, Kling and Seedance generate native audio; Gen-4.5 does not. But native audio does not mean Persian speech, and the next section shows why.

Which models publish numbers today

The AI section of this site already ranks four video models with published weights and sources, and we are not rebuilding that table here; the video model ranking lives where it lives. What an editor should take from it is three lines.

Line one: Seedance 2.5 leads the two columns that matter for narrative, 30 seconds in one generation and 50 reference files, being 30 images, 10 videos and 10 audio clips. Line two: the same model loses the resolution column outright, because ByteDance own API documentation allows only 480p and 720p for it while listing four values up to 4k for its own predecessor. Line three: Kling VIDEO 3.0 is the only model here with native 4K, but its reference-driven mode caps at 1080p, so 4K and heavy referencing do not happen in one generation.

On the day this lesson was written we re-read three numbers from vendor pages that are not in that table. First, the Veo 3.1 documentation now publishes a latency range: minimum 11 seconds, maximum 6 minutes at peak hours. Second, the same page says a generated video is stored on the server for two days and then removed. Third, the Sora 2 model page opens on the OpenAI developer domain and prints a price per second: $0.10 for Sora 2 and $0.30 for Sora 2 Pro, at 720x1280 or 1280x720. That page does not state a clip length for Sora, and neither do we.

The other prices come from the Runway API price list, where one credit is one cent: Gen-4.5 at twelve credits per second, Seedance 2.5 at 720p at thirty, and Veo 3.1 with audio at forty. That is twelve, thirty and forty cents per second of output. Kling is not on that list and its own pricing is in credits tied to a plan, so we leave its cell empty rather than filling it.

The last column is there on purpose. Three of the five prices are a reseller price rather than the maker own, and that difference matters the moment you have to set a budget. It is also why the table in the AI section carries no price column at all, and we have solved that here with a source column instead of by hiding it.

What these are good for today, and what they are not

Our position is plain and fits in two lines. For ideas, mood and filler shots, these tools work today and are cheaper than any alternative. For a brand final, they do not, and anyone claiming otherwise has either never delivered a final or never met a strict brand.

Three uses that genuinely work: the concept shot that sells an idea to a client before real production, the filler or background shot inside a video whose backbone is filmed footage or motion graphics, and the style frame, a generated frame that tells the team what the light, the colour and the texture are supposed to be. In all three the model output is an input to the next stage, not the delivery.

And three places where the answer is no today: any shot in which a real person or a real product has to be exactly itself, any video in which a character must hold across more than a couple of shots, and any output that has to speak and speaks Persian.

Let us source that third one, because it is where Persian tutorials make the most confident wrong claim. The Veo documentation carries a sentence that settles it: English is fully supported, other languages have not been evaluated, so they may work but results can vary. The Kling guide, for its part, counts five speech languages, Chinese, English, Japanese, Korean and Spanish, and Persian is not among them. So for a Persian video, whichever route you take, the voice is a separate step. If you read somewhere that one of these speaks Persian, ask for the source.

There is also a practical constraint few people weigh before the first invoice: cost scales with seconds rather than with attempts, but attempts are seconds too. One 8 second Veo shot at forty cents a second is about three dollars twenty, and if it takes five generations to land the shot you want, that single shot cost sixteen dollars. Nobody says that number in a promotional video.

What leaves with the file

Two things, and neither shows up in the preview.

First, a mark. The Veo documentation states plainly that videos created by Veo are watermarked using SynthID and can be verified on the SynthID verification platform. Which means the thing we know from images is in video too: a mark invisible to the eye that its maker can read. If you want to know exactly what sits inside the file, who can read it and what survives an edit, the lesson on image watermarks and content credentials opens that and covers both media.

Second, an expiry date almost no tutorial mentions. The Veo page says a generated video is stored on the server for two days and then removed, and that to keep a local copy you must download it within those two days. An extended video counts as a newly generated one and its clock restarts. For anyone spreading a project across several days, that means the output of day one may be gone by day four.

A third constraint is worth naming, not as a risk but as a cost: the latency range. The same page gives a minimum of 11 seconds and a maximum of 6 minutes at peak hours. If your plan is to generate and choose between ten shots in one sitting, it is the maximum that sets the length of that sitting, not the minimum. The other three vendors publish no generation time at all, and we invent none for them.

Decide these four things before the first generation

All four get expensive or impossible after generation.

Aspect ratio. Portrait or landscape, from the start. Cropping landscape into portrait means throwing away both sides of the frame, and in a shot where the subject is off-centre that crop cuts the subject. Veo takes the ratio in the request itself, so changing it afterwards means generating again, which means paying again.

The length of each shot, against your own model ceiling. If your model gives 8 seconds, write the shot list in shots of 8 seconds and less, not in twelve second shots you will have to break later. Breaking a shot after generation leaves you with an unwanted cut in the middle of it.

Where the sound comes from. From the model, or separately. If you have Persian narration the answer is decided in advance: separately. In that case it is better to generate the shots silent and build the whole audio bed in one pass, because the model native audio rarely sits under Persian narration and pulling it back out of a video file is hard work. The lesson on AI voice and dubbing opens that stage properly.

What has to stay identical across every shot. A short list: the product colour, the character clothing, the direction of the light, the palette. That list later gets pasted verbatim at the head of every prompt, and it is what stands in for the memory the model does not have. The fast path below builds exactly that.

And a fifth decision that is not technical: whether you tell the client that part of the video was generated. The right answer is that you tell them, and better before you start than after you deliver. For work that is going to be delivered, our motion graphics service settles that boundary with the client up front.