How to Use Seedance: A Beginner's Guide to First Clips

Learning how to use Seedance takes about four decisions and one text field. ByteDance's video model turns a written prompt, plus any reference images you attach, into a finished clip with sound already mixed in. The part that catches people out is not the prompt. It is that duration, resolution, aspect ratio and audio are settings rather than things you describe, and that the bill is counted per second of output.

Start by working out which version you have

Two versions are in circulation and they behave differently. Seedance 2.5, announced 31 July 2026, generates 4 to 30 seconds per request. Seedance 2.0 stops at 15.

A free test settles it. Request a 20-second generation. If it comes back rejected on the duration field you are on 2.0, and because the rejection happens before a task is created, nothing is charged. The shipped 2.5 build is also identifiable by a model string ending in -face.

Everything below applies to both versions except the numbers that carry a version label.

How to use Seedance: the four settings

These are fields. Writing "make it vertical and about ten seconds" in the prompt does not set them.

Choose the aspect ratio from where the clip will be watched rather than from what looks good in a preview. Choose the duration from the number of actions in the scene: five seconds holds one, and a sequence of four things happening needs twenty or more.

There is no 4K option. Output stops at 1080p, and any guide that lists a 4K tier for this model is repeating pre-release material that never shipped.

Write the prompt, dialogue included

The prompt field accepts up to 20,000 characters. It is required when you are generating from text alone, and optional once reference images are attached, where it steers them instead of describing everything.

Describe the subject, the action, the setting, the light and the camera. Keep each of those checkable against the finished clip so a second attempt can be aimed at whichever one came back wrong.

Dialogue goes in quotation marks. Whatever is between the quotes is generated as speech, in the same pass as the picture, so read the line aloud and time it against the duration you set. Sound that is not dialogue gets described rather than quoted.

Reference images are the quickest quality gain available to a beginner. A request takes up to 30 of them, they cost nothing extra, and one photograph of the room you mean settles more than a paragraph about it. Reference clips and audio tracks are supported too, up to 10 each on 2.5, though reference video is the one input that adds to the bill. All media is passed as a public HTTP(S) URL; data URIs are rejected.

Submit, then wait

Generation is asynchronous. The request creates a task, and the video is collected afterwards by polling for that task. Nothing streams back in real time.

Expect two to five minutes on average. Polling costs nothing, so a check every fifteen or twenty seconds is fine. Failed jobs are refunded in full, which includes jobs stopped by content moderation.

Keep the bill predictable

Billing is per second of finished video at a rate set by the resolution. For generation without a reference clip, the per-second rates are $0.118589 at 480p, $0.266824 at 720p and $0.461856 at 1080p.

In whole clips, that comes out as roughly $0.59 for five seconds at 480p, just over $1.33 for five seconds at 720p, and about $4.62 for ten seconds at 1080p.

The habit that saves the most money is drafting low. Iterate at 480p using the exact duration and aspect ratio you plan to ship, because those affect composition and pacing, then run the finished prompt once at the resolution you actually need. Six ten-second attempts at 480p come to about $7.12, and the single 1080p run that follows is about $4.62. Six ten-second attempts at 1080p come to about $27.71.

What usually goes wrong first

Dialogue that is too long for the duration is the most common disappointment, and it is arithmetic rather than a model failure. Count the words.

The second is settings written as prose. Aspect ratio, duration and resolution are ignored when they appear in the prompt text.

The third is a rejected media URL. The file has to be reachable over public HTTP(S), and inline data is refused across the platform.

The fourth is error 80006, a safety refusal that does not say which stage fired. Named real celebrities, third-party intellectual property and illegal content are blocked inside the model and no setting removes them. A content_filter field exists and defaults to true; setting it false selects a less restrictive channel and leaves moderation in place. The job is refunded either way.

Frequently asked questions

Do I need to know how to code to use Seedance? Not for a hosted interface, which exposes the same duration, resolution, aspect and prompt fields as a form. Calling the hosted API directly needs an API key and the ability to make an HTTP request and poll for a result.

How long does a Seedance video take to generate? Two to five minutes on average. The job is asynchronous, so the request returns a task immediately and the finished file is collected by polling.

What is the cheapest way to learn how to use Seedance? Draft at 480p. It is the lowest per-second rate, generation time is the same, polling is free, and failed jobs are refunded, so experimentation costs cents rather than dollars.

Can I add my own images or voice? Reference images and audio tracks are accepted and add nothing to the bill. Reference video is accepted too, and it is the one input type that increases the billed seconds.

The short version

How to use Seedance, condensed: confirm the version, set duration, resolution, aspect ratio and audio as fields, write a prompt that describes the shot and quotes the dialogue, attach reference images because they are free, submit and poll for two to five minutes, and do all the iterating at 480p before paying for one 1080p run.