← All articlesTools

Seedance 2.0 and 2.5: a guide to multi-shot, references and access

How Seedance 2.0 and 2.5 handle multi-shot and references, how to write a timed prompt, where to use it in November 2026, and what the Hollywood dispute teaches.

Seedance 2.0 is ByteDance’s video model, released in February 2026: in a single pass it generates 4 to 15 seconds of multi-shot video with native audio and lip sync, from text and references of up to 9 images, 3 videos and 3 audio tracks. On 31 July 2026 Seedance 2.5 followed, taking a single take to 30 seconds and references to 50. As of November 2026 it is mostly used through third-party platforms, with limits that vary from country to country.

It is the engine that has caused the most argument this year, over both its quality and its rights record. Here is how multi-shot works, how to write references, where you can use it and what to weigh up before putting it in a campaign.

What Seedance 2.0 is and what changes with 2.5

Seedance is ByteDance’s family of text-to-video and image-to-video models, first released in June 2025 according to Wikipedia. Version 2.0 went viral within days of its release in February 2026, for its realism, its motion and camera control and its 15-second clips. A smaller variant, Seedance 2.0 mini, arrived in June; version 2.5 at the end of July.

FeatureSeedance 2.0Seedance 2.5
ReleasedFebruary 202631 July 2026
Length4-15 seconds4-30 seconds
References9 images, 3 videos, 3 audio30 images, 10 videos, 10 audio
Resolution on fal480p, 720p480p, 720p, 1080p
FormatsSix, from 21:9 to 9:16Seven, from 21:9 to 9:16
StructureSeveral shots with cutsSingle take, area editing

The figures come from fal’s pages for Seedance 2.0 and Seedance 2.5; the latter was updated on 21 October 2026. For 2.5, Wikipedia’s text-to-video entry adds 4K output and a long-video mode, in beta, that extends clips to 3 minutes. There is also Seedance 2.0 Fast, quicker and cheaper, for test runs.

How Seedance multi-shot works

Seedance 2.0's multi-shot produces several shots, with cuts and transitions, inside a single 15-second output. With 2.5 the logic changes: the model generates one continuous take, and shot changes are written into the prompt as a timed shot list. According to fal, the time ranges must run consecutively without overlapping, and they guide the pacing rather than marking frame-accurate cuts.

In practice the prompt becomes a shooting script. Here is how one might begin for a 15-second running-shoe teaser:

0-4 s: wide shot, athletics track at dawn, @runner enters from the left, lateral tracking shot. 4-9 s: medium-long shot, running towards camera, breathing and footsteps up front in the mix. 9-13 s: close-up of @sneaker on wet tarmac, low camera. 13-15 s: front-on packshot, no camera movement, ambience only.

Writing it well takes a few steps, and they are always the same:

  1. Open with context: location, time of day, light, visual style.
  2. List the active references and what each one is for.
  3. Split the length into time ranges, one for each beat of the story.
  4. For each range, state the shot size, camera move, action and sound.
  5. Close with constraints: no slow motion, no split screen, no unplanned cuts.

It is the same work as a storyboard, just written in another form. Skip it, and the model decides for you.

References: how to use them without losing control

Seedance’s strength is how much material it accepts: with 2.5, up to 50 references, with video and audio up to 30 seconds in total. More material does not mean more control unless it is organised. A workflow published by Higgsfield shows a method that works.

  • One reference, one job. Separate images for face, costume, location and scale, called up with a tag such as @location or @size-ref.
  • Say what to ignore. If a reference is only there for the face, tell the model to ignore its background and lighting.
  • Respect the source. Video-to-video needs at least 4 seconds of source footage, and an output longer than the source pushes the model to improvise.
  • Read repeated errors. If the same flaw appears in four generations out of four, the problem is the prompt or the source, not luck.

With 2.5 you can also have a character sing or speak over an audio track you supply, with lip sync to match. For an artist, that means starting from the real song rather than a generated voice: more on this in our guide to the AI music video.

Where to use Seedance 2.0 in November 2026

For a European brand, Seedance almost always comes through a third-party platform, and each has its own rules.

  • fal. API access since 9 April 2026 for 2.0, later for 2.5. For US-hosted inference, faces and IP permissions go through a separate approval process.
  • Higgsfield. Offers 2.0, billed as native 4K, and 2.5, whose page mentions both native 1080p and native 4K.
  • Adobe Firefly. Includes Dreamina Seedance 2.0, but Adobe’s partner models page limits it to individual plans, and an Adobe help article from July 2026 states that it is not available to users in the United States.

So the same model comes out at 720p on one platform and in 4K on another. Before you promise a client a resolution, check it on the platform you will actually use. Kult Studios is a Higgsfield and Adobe partner; bear that in mind.

Rights: what the Hollywood case teaches

Within days of 2.0's release, clips featuring famous actors and characters from films and series were circulating. According to Wikipedia’s account, on 13 February 2026 the Motion Picture Association accused the model of copyright infringement and Disney sent ByteDance a cease-and-desist letter. Paramount’s followed on 15 February. On 16 February ByteDance said it “respects intellectual property rights” and would strengthen its safeguards. In March two US senators called for it to be shut down.

For a campaign the consequences are practical. Other people’s characters, actors and brands are not used as references. Real faces come in only with written consent. Usage rights for the output are settled before generation, together with the terms of the chosen platform. The wider picture of ongoing cases is in our piece on AI copyright lawsuits.

Seedance 2.0 and 2.5: strengths and limits

The strength is multi-shot storytelling with synchronised audio. In a comparison published by Higgsfield, one generation per category with no retouching, 2.5 held lip sync even through a scream and landed its cuts on the beat, while 2.0 lost lip sync in close-ups. It is a test by a platform that sells the model, but it shows clearly where the step up lies.

The limits need factoring in from the start:

  • Languages. According to fal, 2.5 generates dialogue in eleven languages, including English, Spanish, Portuguese, Japanese and Arabic. Italian, French and German are not among them.
  • Timing precision. Timestamps guide the pacing, not frame-accurate cuts: fine editing stays in post.
  • Availability. Plans, countries and resolutions vary by platform, and may change again.
  • Reputation. For some clients, the model’s legal history weighs as much as its image quality.

Multi-shot does not replace the storyboard

Seedance generates cuts, but it does not decide which ones the story needs. That is why, in our method, shots are written and approved in pre-production, and only then is the engine chosen, whether Seedance or another of the available AI video models.

If you are working on a music video or a launch and want to know whether Seedance is the right engine, see how we approach music videos.

Frequently asked questions

What is the difference between Seedance 2.0 and Seedance 2.5?

Seedance 2.0 generates up to 15 seconds across several shots and accepts up to 15 references. Seedance 2.5, released on 31 July 2026, reaches 30 seconds in a single take, accepts up to 50 references and can edit single areas of the frame.

Can I use Seedance 2.0 in the US?

Through fal, yes, with a separate approval process for faces and IP permissions on US-hosted inference. In Adobe Firefly, however, Adobe stated in July 2026 that Seedance models were not available to users in the United States.

Which languages does Seedance 2.5 speak?

According to fal’s documentation, Seedance 2.5 generates dialogue natively in eleven languages, including Chinese, English, Spanish, Portuguese, Japanese, Korean and Arabic. French, German and Italian are not on the list.

Sources

  1. Seedance · Wikipedia, 2026-11
  2. Seedance 2.0 · Wikipedia, 2026-11
  3. Text-to-video · Wikipedia, 2026-11
  4. Seedance 2.0 API Live on fal · fal, 2026-04
  5. Seedance 2.5 on fal · fal, 2026-10-21
  6. Seedance 2.0 on Higgsfield · Higgsfield, 2026-11
  7. Seedance 2.5 on Higgsfield · Higgsfield, 2026-11
  8. Seedance 2.5 vs Seedance 2.0: We Tested Both in 6 Categories · Higgsfield, 2026
  9. How To Make Realistic VFX Shots With Seedance 2.5 (Full Workflow) · Higgsfield, 2026
  10. AI Video Generator: Kling, Veo, Seedance and more · Higgsfield, 2026-11
  11. Partner models in Adobe products · Adobe Help Center, 2026-10-23
  12. Generate videos using Seedance · Adobe Help Center, 2026-07-27

Got an idea? Let’s make it a film.

We’ll reply with a concrete proposal, built around your brief.