It is a video generation model developed by ByteDance, the Chinese company behind TikTok and CapCut, in its Seed team, which develops the company's AI models. It is the engine that follows long prompts with several chained actions best: where other models lose a piece of the scene along the way, Seedance remembers them all. The current version, Seedance 2.5, reaches thirty seconds in a single generation and accepts up to fifty references across images, video and audio (at most thirty images).

You can try it with the free credits of Dreamina, CapCut's creative platform, depending on whether the model is available in your country and plan. For serious production there is no Seedance subscription for the general public: you use the engine from ByteDance's apps such as Dreamina, from the platforms that host it, Higgsfield included, where it is included in their plans, or, for companies, from BytePlus's API.

Who created Seedance?

ByteDance, the Chinese company behind TikTok and CapCut, with its Seed team, which develops the company's AI models. Seedance was born as a video model in 2025 and has been updated several times since, up to version 2.5.

How long does a video generated with Seedance last?

With Seedance 2.5 a generation can reach thirty seconds: a jump compared to the few seconds typical of video models. In practice, production clips often stay shorter, because the longer the scene the more ways the model has to get it wrong, and the successful ones get selected and edited.

Seedance or Veo: which one to choose?

It depends on the scene. If the scene has several actions in a row, a character entering, picking something up and leaving, Seedance holds the sequence together. If the scene is a single shot written to the millimeter, with the camera specified, Veo 3.1 executes it with more precision. People who write whole scenes love Seedance; people who write shots love Veo.

Where do you use Seedance?

You try it in Dreamina, CapCut's creative platform, which gives free credits to start (Seedance 2.5 depends on your plan and country). To produce, you use it inside the platforms that host it in their plans: I use it mostly in Higgsfield, together with the directing tools that surround the engines.

Does Seedance work in Italian?

It accepts prompts in Italian, but like all video models it works better with English: my prompts always start in English even when the final video is for the Italian audience. On long prompts, English also gives an extra advantage: the model follows them with more precision piece by piece.

← Blog Guides

Seedance: what it is, how to use it and how you pay

Seedance by ByteDance explained by someone who uses it: the engine that remembers multi-action scenes, Seedance 2.5's 30 seconds, where to try it, how to pay.

Erika Taranto

Official selection · AI for Good 2026 10 min read

Published September 4, 2026 Updated October 25, 2026

Seedance: what it is, how to use it and how you pay

In short

Seedance is the video generation model of ByteDance, the company behind TikTok and CapCut, developed by its Seed team. Among the five big engines it is the one that follows long prompts with several chained actions best: where the others lose a piece of the scene along the way, Seedance remembers them all. The current version, Seedance 2.5, reaches thirty seconds in a single generation and accepts up to fifty references across images, video and audio. You can try it with the free credits of Dreamina, CapCut's creative platform, depending on whether the model is available in your country and plan, and you use it in production inside the platforms that host it: I use it mostly in Higgsfield.

All video models can shoot a shot. Few can shoot a scene. The difference is this: the shot is a woman at a cafe table; the scene is the woman entering, sitting down, ordering and opening the newspaper, in this order, without the order getting lost. Seedance is the engine I call the interpreter, because it reads the scene whole and performs it whole: the others, when the scene has three things that must happen in sequence, sooner or later forget one.

It is also the engine I use most inside the platform where I work every day. This guide zooms in on it: the complete map of the five engines is in create videos with AI.

Everyone can shoot the shot. Not the scene: for that you need an engine that remembers what comes next.

What Seedance is (and who owns it)

Seedance is a text-to-video and image-to-video model developed by ByteDance, the Chinese giant behind TikTok and CapCut, in its Seed team, the one that develops the company's AI models. Born as a video model in 2025, it has been updated at a fast pace: version 2.5, released in late July 2026, is the one I talk about here.

Seedance has a clear place in my toolbox: it is the interpreter of texts. When the scene is written well, with several actions following each other, Seedance reads it all and returns it all. If you write scenes, not single moving images, you will love it: it is the engine that treats your prompt as a script instead of a suggestion.

The problem it solves: the piece lost along the way

The flaw all of us who make AI videos know: you ask for three actions and the model returns two, or mixes them. "Enter, sit down, open the letter" becomes "enter, open the letter", and the chair disappears. On engines that think in shots it is almost built in: the model favors the best-looking scene, not the requested sequence.

Seedance tackles the problem at the root, and in my weeks of using it, it holds up: chained actions stay in the written order. That is why narrative scenes, where one gesture sets up the next, run through it: in an ad where the hand picks up the product, turns it and sets it down, the order is the message.

Two sequences compared: the other models lose an action along the way, Seedance keeps all three in the written order — Three actions in a row: where the other models lose a piece, Seedance keeps the sequence.

Seedance 2.5: thirty seconds and fifty references

Version 2.5 made two leaps that matter, verified in October 2026:

Thirty seconds in a single generation. The other engines work in cuts: clips of a few seconds that then get edited. Seedance 2.5 can hold a whole scene in a single thirty-second shot, with all that means for storytelling: the scene breathes without a cut.
Up to fifty references. You can give the model a whole album of starting points (up to thirty images, plus video and audio): the character, the environment, the objects, and the video respects them. It is the function that changes the work on projects with brands and recurring characters, where consistency is the deliverable.

If the reference is the face of a real person, platforms may ask for verification and authorization to use the likeness: check the rules of the one you use.

As for the thirty seconds, an honest word from the trade: the longer the scene, the more ways the model has to get it wrong. The long take is one more option, not the only route: the best results still often come from medium durations, and you choose shot by shot.

The two numbers of Seedance 2.5: thirty seconds in a single generation and up to fifty references at the start — The two numbers of 2.5: thirty seconds in one generation, fifty references at the start.

Where you try it

Seedance has no subscription of its own for the general public: it is an engine, and engines live inside apps and platforms (for companies there is also BytePlus's API). Two routes to start:

Dreamina, CapCut's creative platform: it gives free trial credits to eligible users, but Seedance 2.5 may require a subscription and depends on the country. Open the account and look at the model selector: if it is there, write the scene and see what kind of interpreter it is.
Higgsfield and the platforms that host it: for serious production I log in through higgsfield.ai, where Seedance is included in the plans together with the other engines and the directing tools. The complete guide to the platform is the article Higgsfield: what it is, how to use it and how you pay.

Update of October 16, 2026. At launch, on July 31, 2026, Dreamina announced Seedance 2.5 for subscriber accounts in Europe, Asia, the Middle East and South America; on August 6 BytePlus announced the model in its ModelArk API too, in supported markets (BytePlus is not available in the United States). A free trial of 2.5 is not confirmed by the official sources.

The difference between the two routes. Dreamina is for trying the engine; the platform is for working with it. If you need one clip, the first is enough; if every week you need ten, with character consistency and formats ready for social, the second pays for itself in saved regenerations. The complete picture of the free routes is in free AI video generators.

How to use it: the first scene in five steps

Open Dreamina and choose video generation. Free account, no card to start; Seedance 2.5 may require a subscription, so check that it is available in your plan and country.
Write the scene like a script, not like a list. Below is the exact format I use, with the actions in order.
Upload the references if you have them. The character, the environment, the brand's object: on Seedance, references are not an oracle: they are concrete starting points the video respects. If it is the face of a real person, check the platform's rules.
Choose the duration with judgment: try the scene on a medium duration, and only when the script holds ask for the long take. Thirty seconds of mistakes cost as much as thirty seconds that work.
When you produce for work, move to the platform: in Higgsfield Seedance stands next to Kling and the directing tools, and you can manage a flow of several clips with the same consistency without jumping between four sites.

Seedance: ByteDance's video engine that keeps three actions in a row.

The script prompt: how to write for Seedance

The format I use, built around the engine's strength. The actions in order, connected, and the camera declared once:

A baker in a white apron enters the workshop, turns on the oven, kneads the dough on the wooden table; camera holds on a wide shot, warm light from the oven growing as it heats.

Note how the three actions sit in one sentence: enters, turns on, kneads, in this order and without periods and line breaks the model could read as separate scenes. The camera sits at the end, declared once: wide shot, still. The light has a logic that evolves with the action (the oven lighting up as it heats): it is the kind of detail Seedance picks up and uses.

The general rules of the video prompt, including the traps, are in the dedicated section of the main guide: one sentence, five elements, no negations. On Seedance you add the script rule: actions in order, tied together in a single sentence.

The anatomy of the script prompt for Seedance: subject, three actions in order, camera declared once, light following the action — The script prompt: actions in order, camera declared once, light following the action.

How much Seedance costs

Seedance has no price list of its own: the price depends on the platform it runs on. The overall picture, verified in October 2026:

Route	What you pay	When it makes sense
Dreamina (CapCut)	free credits to start, then Dreamina's plans (Seedance 2.5 may require a subscription)	Trying it out, learning to write scripts
Platforms that host it	included in the plans: in Higgsfield the clip is paid in plan credits, and the cost grows with duration and resolution	Producing every week
No direct subscription	there is no "Seedance plan" for everyone: the engine is paid through the app or platform that hosts it, or through BytePlus's API for companies	Worth knowing before you search for it

The comparison with the plans of Veo and Kling is in the costs section of the main guide. And the rule always holds: the real cost is the attempts, and a long scene that goes wrong is paid for second by second.

Where Seedance lives: Dreamina to try it, the platforms that host it to produce, the plan credits in Higgsfield — Where it lives and what it costs: Dreamina to try, the platforms' plans to produce. Overview verified in October 2026.

The honest limits

Body physics. It holds the sequences, but when a body has to look alive in a difficult movement, like a dance step or a run, I send it to Kling: that remains its territory.
Directing precision. If the shot is written to the millimeter, Veo 3.1 executes with more fidelity: Seedance favors the sequence, Veo the description.
The long take carries a growing risk. Thirty seconds are an option, not a promise: the longer the generation, the more corrections cost. In practice I alternate durations.
It has no home of its own. Search for "the Seedance website" and you find the Seed team's blog, not an app: to use it you go through Dreamina or a platform. It is not a flaw, but you need to know it.
English pays off twice. It accepts Italian, but script prompts work better in English, where the model distinguishes the sequence of verbs better.

Long take or editing: the single thirty-second strip against the pieces to choose from and edit — Long take or clip editing: when thirty seconds are worth it and when not.

When I choose Seedance, and when I don’t

Seedance gets the scenes written like scripts: several actions, sequences, a long take when needed. I give the rest to other tools like this:

the shot to the millimeter: Veo 3.1;
the living body, gestures and walks: Kling;
whole projects with consistency, camera and social formats: Higgsfield, where Seedance is the engine I use most.

And if the scene you have in mind is for your brand and you would rather work with someone who writes scenes every day, that is my craft. Write to me and we start from an idea.

The guides on the video path

Create videos with AI, the map of all the tools (the pillar of this guide);
Higgsfield, the platform where Seedance works with the directing around it;
Higgsfield for Blender, where Seedance executes to the letter the movement you prepare in 3D;
Veo 3.1, the engine that executes the written scene;
Kling AI, the realist;
Sora 2, the storyboard pioneer, which closed in 2026;
Free AI video generators, all the ways at zero cost.

Sources

ByteDance Seed, Seedance 2.5, official announcement: seed.bytedance.com
Dreamina, CapCut's creative platform where you can try the engine: dreamina.capcut.com
Dreamina, launch announcement of Seedance 2.5: dreamina.capcut.com
BytePlus, Seedance 2.5 on ModelArk, access and rules on faces: byteplus.com
Higgsfield, the platform that hosts it in the plans, Seedance 2.5 on Higgsfield: higgsfield.ai
Maximum duration and references: verified in October 2026, same check as the main guide