How to Use Seedance 2.0: Settings, Limits and Cost

Seedance 2.0 is the older of ByteDance's two current video models, and it is still what many hosts serve by default. Learning how to use Seedance 2.0 is mostly a matter of working inside one hard limit: a single generation stops at 15 seconds. Resolution, prompt, reference material and cost all follow from what you decide to do with that window.

This page covers the mechanics. What to set, in what order, and what each choice does to the bill.

Decide the duration before anything else

Fifteen seconds is architectural, not a plan restriction. No account tier, no host and no parameter combination extends it. If your finished piece is 40 seconds, you are producing at least three generations and joining them in an editor, and the joins are what will look wrong, so plan the cuts before you write a prompt.

Shorter is also cheaper in a direct, linear way. Billing is per second of finished video, so a 6-second clip costs roughly 40 percent of a 15-second one at the same resolution. The maximum is more than a single shot usually needs. One action, one camera move, one line of dialogue tends to fit in 5 to 8 seconds, and asking for 15 gives the model room to drift into a second idea you did not want.

Pick a resolution that matches the delivery, not the ambition

Three tiers exist: 480p, 720p and 1080p. There is no 4K option in this model family, on any host, at any price. Pages that list a 4K tier are quoting pre-release material.

The practical approach is to iterate at 480p and finalise at the resolution you actually ship. Composition, timing, camera behaviour and whether the model understood the prompt are all legible at 480p, and the per-second rate there is a fraction of the 1080p rate. Once the shot is right, rerun it at 1080p with the same prompt and seed. The only thing you cannot judge at 480p is fine texture: skin, fabric weave, small text on signage.

Write the prompt as a shot description

The model responds to the same information a camera operator would need. Subject, what the subject does, where the camera is, how the camera moves, what the light is doing. Vague quality adjectives do less than a concrete noun.

Two habits that improve results on this version specifically:

Give one action per clip. A prompt describing three consecutive events tends to produce a compressed, rushed version of all three inside the same 15 seconds. Describe one, generate it, then generate the next.

State the ending. Where the subject is and where the camera is at the final frame. Without it, the last second often drifts.

Audio generation is part of the model rather than a separate pass, so dialogue, effects and ambience are produced together with the picture. If you intend to replace the audio in an edit anyway, turn it off and save yourself the sync problems.

Use reference images to lock what words cannot

Text is poor at specifying a face, a product, a logo or an exact colour. A reference image fixes those directly, and reference images add nothing to the bill. Reference video is the input that costs money, because the platform charges for the source clip's own length on top of the output.

Seedance 2.0 accepts a narrower set of reference material than 2.5 does, and it does not edit or extend existing footage. Those capabilities arrived with the newer version. On 2.0 the reference is guidance for a fresh generation rather than a source you are modifying.

All media inputs must be public HTTP or HTTPS URLs. Data URIs are rejected across the platform, so a local file needs to be uploaded somewhere reachable first.

How to use Seedance 2.0 without burning money on failures

Generation is asynchronous. You submit a job, get a task id, and poll until it finishes. Polling is free, so there is no reason to poll slowly. Typical completion is two to five minutes, and a job that has been running for twenty is a job that has gone wrong.

Failed jobs are refunded in full. That includes refusals from content moderation, which is the failure mode most people hit first.

Moderation sits in two places. The content_filter parameter defaults to true; setting it to false routes the request to a less restrictive channel, and it does not turn moderation off. Underneath that, the model itself blocks named real celebrities, recognisable third-party intellectual property, and illegal content, and no host setting removes those.

A refusal returns error 80006. It does not tell you which stage fired: the prompt, a reference image, or the generated output. Diagnosis is manual. Strip the prompt back to a neutral description, confirm it passes, then add elements back one at a time until it fails again.

What a clip actually costs

Rates are per second of finished video and depend on resolution. The published figures for Seedance 2.5 are below, and they function as a ceiling for 2.0, which is priced lower per second.

Two billing rules are worth knowing before you send a request containing video. A reference clip is charged for its own length in addition to the output, and any request carrying video has a floor of five thirds of the output duration. A two-second source with a five-second output bills nine seconds rather than seven.

Because 2.0 caps at 15 seconds, its worst case per generation is bounded in a way 2.5's is not. That predictability is the main reason to stay on the older version for high-volume work.

Frequently asked questions

What is the longest clip Seedance 2.0 can make? Fifteen seconds in a single generation. Longer pieces are separate generations joined in an editor, or a move to Seedance 2.5, which runs to 30 seconds.

Can Seedance 2.0 output 4K? No. The ladder is 480p, 720p and 1080p. There is no 4K tier anywhere in this model family.

Why was my prompt refused when it contained nothing unusual? Error 80006 covers every safety stage without naming the one that fired, and reference images are checked as well as text. Isolate by simplifying the prompt, verifying it passes, then reintroducing elements individually. The failed job is refunded.

Does turning off the content filter give uncensored output? No. Setting content_filter to false selects a less restrictive channel. Moderation still applies, and the model-level blocks on real celebrities, third-party IP and illegal content stay in place regardless.

The short version

How to use Seedance 2.0 comes down to four decisions. Keep the clip short, because 15 seconds is a wall and every second is billed. Iterate at 480p and finalise at delivery resolution. Describe one action and name the final frame. Use reference images freely, since they are free, and treat reference video as a paid input with its own billing floor. Failures cost nothing, so the expensive mistake is not a refusal but a 1080p render of a prompt you had not tested at 480p.