What Is Seedance? ByteDance's Video Model Explained
What is Seedance: ByteDance's video generation model family, what the 2.0 and 2.5 versions each produce, what they accept as input, and what a clip costs.
Seedance is a family of video generation models made by ByteDance. Give it text, still images, audio or existing footage and it returns a finished video clip with sound already on it. Anyone asking what is Seedance usually arrives with a second question attached, which is which version they are actually getting, because two are in circulation and their limits differ enough to change what you can make.
What is Seedance, in one paragraph
Seedance takes a written prompt and optional reference media, and produces a single video file with a synchronized audio track. Speech, sound effects and music are generated in the same pass as the picture rather than added afterwards, and the audio comes back mono. A generation can contain more than one shot, so a request is closer to a short scene than a single continuous take.
Generation is asynchronous. A request creates a task, the task takes an average of two to five minutes, and the result is collected by polling. Polling itself is free, and a job that fails is refunded in full.
The two versions and what separates them
| Seedance 2.0 | Seedance 2.5 | |
|---|---|---|
| Maximum clip length | 15 seconds | 30 seconds |
| Reference media | narrower | 30 images, 10 clips, 10 audio tracks |
| Edit or extend existing footage | no | yes |
| Per-second price | lower | higher |
Seedance 2.5 was announced by ByteDance on 31 July 2026. The clip ceiling is the difference that decides most projects: 30 seconds in one pass removes a stitching step for anything longer than a short cut, while 15 seconds forces segments and continuity work.
Native audio, multi-shot scenes and realistic motion are common to both. Moving between them is largely a matter of changing the model string, since the request shape is close.
What Seedance accepts as input
The prompt field takes up to 20,000 characters. It is required for text-to-video and optional when reference media is supplied, where it directs the references rather than describing everything from scratch. Lines written inside quotation marks come back as spoken dialogue.
On Seedance 2.5 one request can carry up to 30 reference images, up to 10 reference clips of 2 to 30 seconds each with 30 seconds combined, and up to 10 audio tracks. Reference images and audio add nothing to the bill. Reference video is billed.
All media is passed as public HTTP(S) URLs. Data URIs are rejected.
Output settings are chosen by field rather than described in prose:
- Duration: 4 to 30 seconds on 2.5, default 5. Up to 15 on 2.0.
- Resolution: 480p, 720p or 1080p, default 720p.
- Aspect ratio: 16:9, 9:16, 1:1, 4:3, 3:4, 21:9 or adaptive, which is the default.
What Seedance costs
Billing is per second of finished video. The rate depends on the resolution and on whether the request carried a reference clip.
| Resolution | Generation only | With a reference video |
|---|---|---|
| 480p | $0.118589 | $0.071153 |
| 720p | $0.266824 | $0.160094 |
| 1080p | $0.461856 | $0.275904 |
A ten-second 1080p clip generated from scratch is about $4.62. A thirty-second 480p clip is about $3.56.
Two rules offset that lower reference-video rate. A reference clip is billed for its own length on top of the output, and any request containing reference video carries a floor of five thirds of the output duration, rounded up.
What Seedance does not do
There is no 4K tier. Output stops at 1080p, and that ceiling is the same on every host. Pre-release coverage of 2.5 described 4K output and it did not ship, which is why the claim still turns up in comparison tables.
Three categories are blocked inside the model itself, and no host setting removes them: named real celebrities, third-party intellectual property, and illegal content. A content_filter field exists and defaults to true; setting it false routes the request through a less restrictive channel and does not switch moderation off.
Safety refusals surface as error 80006. The error does not identify which stage fired, so the trigger could be the prompt, a reference file or the generated output. Refused jobs are refunded like any other failure.
Frequently asked questions
Who makes Seedance? ByteDance. Seedance 2.5 was announced on 31 July 2026, and the earlier 2.0 remains available alongside it.
Does Seedance generate sound, or just video? Sound is generated with the picture in a single pass, covering speech, effects and music as one mono track. Quoted lines in the prompt become spoken dialogue.
How long can a Seedance video be? Seedance 2.5 generates 4 to 30 seconds per request, defaulting to 5. Seedance 2.0 stops at 15 seconds. Anything longer than the ceiling is chained segments.
Can Seedance output 4K? No. The available resolutions are 480p, 720p and 1080p. There is no 4K setting, plan or host that produces it.
The short version
What is Seedance, condensed: a ByteDance video model that turns a prompt and optional reference media into a clip with sound baked in. Version 2.5, announced 31 July 2026, generates 4 to 30 seconds and takes wide reference input; version 2.0 stops at 15 seconds and costs less per second. Both top out at 1080p, both bill by the second, and both refund failures in full.
