ByteDance Seedance 2.5 is the video model ByteDance published on 31 July 2026: 30 seconds in one generation, 50 reference items per request, editing aimed by timestamp. This guide opens with where it runs.
What ByteDance Seedance 2.5 is
The launch post calls it a continuation of the audio-video architecture introduced with Seedance 2.0, with the work going into long narrative, reference and editing.
| Model ID | Length | Resolution | Reference input | Format |
|---|---|---|---|---|
doubao-seedance-2-5-260628 | 4~30 s | 480p / 720p (8-bit) / 1080p (10-bit) | 50 (30 images + 10 videos + 10 audio) | mp4, mov |
doubao-seedance-2-0-260128 | 4~15 s | 480p / 720p / 1080p / 4k (10-bit) | 15 (9 + 3 + 3) | mp4 |
Two entries are counter-intuitive: the older generation offers 4k output and the current one does not, so 1080p is the ceiling at 24 fps, and doubao-seedance-1-5-pro-251215 is marked retired on the same list. Only this generation takes audio on its own.
The post names the model’s own weak spots: physical plausibility in complex motion, and stability where many subjects interact. A sibling video line, the Veo hub, keeps a similar spec sheet.
Where you can actually use it
Three entrances are documented, and none is a Seedance website.
- Jimeng AI, web — video generation, then choose Seedance 2.5.
- Doubao Pro — the same path.
- Volcengine Ark —
POST /api/v3/contents/generations/taskswith modeldoubao-seedance-2-5-260628.
The project page at seed.bytedance.com/seedance2_5 is a description, not a product: no price, no sign-up, no FAQ, and no consumer site at all. The Seedance hub collects the model line; the Seedance API guide covers the endpoint and the asynchronous task lifecycle.
Why the top search results are not the official site
No official entrance lives under a seedance.* address, so a page presenting itself as the model’s official website is not one. A site such as seedance.tv names no operator and no ByteDance relationship, gives two clip ceilings for the same model, 4–8 seconds and 4–15 seconds, neither matching the documented 4 to 30, and sells monthly plans, while no official page publishes a plan, a subscription or a free allowance at all. The cost is routing, not morals.
Your first clip, step by step
The first run confirms the loop, not a finished clip.
- Pick an entrance: Jimeng AI or Doubao Pro.
- Choose the model: video generation, then Seedance 2.5.
- Write the prompt first: subject, then shots, then sound.
- Set the frame. Ratio is one of 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 or adaptive; duration runs 4 to 30 seconds, or
-1to let the model pick. - Add reference material only when the clip must match something you own. Fifty items are allowed per request.
- Generate, then extend. The English project page says twice; the Chinese launch post gives no number.
Worked example: a prompt the vendor published
Timestamps, not adjectives
The post pairs seven full prompts with their output clips, the only vendor prompt-and-result pairs anywhere. The pattern is segmented timing: the prompt lists 0-5 s, 6-10 s and 11-20 s segments, so the model is told what happens in each part instead of being asked for a mood. Reference material is addressed inline as image 2, video 1 or white model 1, an untextured scene block that fixes framing before the look is rendered.
16:9 widescreen, cinematic, one continuous take, no cuts
Scene reference: @image 1
0-5 s: what the first segment shows, and what the camera does
6-10 s: how the shot moves on, and who speaks
11-20 s: how the segment closes, and where the camera endsReference material, and the real-face rule
Fifty items is the ceiling, split as 30 images, 10 videos and 10 audio clips; the 2.0 series stops at 15. Reference videos run 2 to 30 seconds each, at least 4 seconds on an edit, at most 10 files and 30 seconds in total, 200 MB each, 24 to 60 fps in.
The rule that blocks the most first attempts: the model rejects an uploaded reference image or video containing a real person’s face. Three official routes exist: reuse a face the same account already generated, take a virtual portrait from the platform library, or supply authorised footage.
Aspect ratio and duration rules that cause most errors
The rule table
Two parameters change meaning with the task, which is why a ratio setting gets rejected.
| Task | What the vendor requires |
|---|---|
| Text to video | No special restriction on ratio or duration |
| First frame, first and last | ratio must be adaptive |
| Reference to video | Same as text to video |
| Video editing | ratio adaptive and duration -1; reference video 4–30 s |
| Video extension | ratio must be adaptive |
The same parameters can be appended to the prompt text instead.
--rs 720p --rt 16:9 --dur 5 --seed 11 --cf false --wm trueWhat it costs
Ark bills by tokens, not by plan. The rate depends on output resolution and on whether the request carries a video input: at 480p and 720p, 70.00 and 42.00 yuan per million tokens; at 1080p, 77.00 and 46.00, a limited-time 72 per cent of list. Offline inference is not supported.
Price is the token rate times usage, and usage is estimated as input plus output video seconds, times output width, height and frame rate, divided by 1024; only successful videos are billed.
No official document gives a per-second price, so any such figure is the reader’s own arithmetic. Rate limits are per account: 600 requests per minute and 10 concurrent tasks on enterprise accounts, 180 and 3 on individual ones.
What the vendor has not published
The launch post states no resolution tier, no frame rate and no reference file-size cap; the first two come from the Ark model list instead. A success rate, a consistency percentage, a benchmark score and a parameter count exist nowhere official either, so a site quoting them is quoting its own estimate.
Frequently asked questions
Is there an official Seedance website? No. The entrances are Jimeng AI, Doubao Pro and Volcengine Ark.
How long can one clip be? Up to 30 seconds, with extension beyond that. The English project page says twice; the Chinese launch post gives no number.
Can I use a real person’s face? Not by uploading it. Reuse a face generated on the same account, take a virtual portrait from the platform library, or use authorised footage.
Is there a free tier? The vendor publishes no free allowance; clips are billed per million tokens at the rates above.
Learn the entrances and the ratio rules once, and the rest stops feeling like a slot machine.