Seedance 2.5 is ByteDance's model for generating video with sound. In one pass it assembles a clip up to 30 seconds long and accepts up to 50 references (images, video, audio). There is nothing to download: it is a model, not a program. You can use it through the Dreamina service or over the API, and codexkey offers it under the ID seedance-2.5.
What Seedance 2.5 is
ByteDance showed the model on July 31, 2026 (TechNode). The Seed page describes it as a model for joint audio and video generation, for stories up to 30 seconds, and a clip can be extended twice more. Among the professional features it lists control of a "white model" and green-screen work (seed.bytedance.com).
At first the model was rolled out in the Chinese apps Jimeng AI and Doubao. The API for international customers arrived later, through BytePlus.
The BytePlus documentation lists these modes: text to video, an image as the first frame, a "first and last frame" pair, omni-reference (up to 30 images, 10 videos and 10 audio clips plus text), editing an existing video, and extending it. For editing, extending and first-frame mode, the frame format comes from the source material. In the modes with references, you choose the length and format.
Seedance 2.5 vs 2.0: what changed
The figures below come from the BytePlus ModelArk documentation as of October 29, 2026.
| Parameter | Seedance 2.0 | Seedance 2.5 |
|---|---|---|
| Length per pass | 4-15 s | up to 30 s |
| Reference images | up to 9 | up to 30 |
| Reference videos | up to 3, 15 s in total | up to 10, 30 s in total |
| Reference audio | up to 3 | up to 10, a single audio clip is accepted as input |
| Resolution | 480p, 720p, 1080p, 4K | 480p, 720p, 1080p |
| Video editing and extension | yes | yes |
The takeaway is not the obvious one: the documentation leaves 4K to Seedance 2.0. Version 2.5 wins on clip length, on the number of references and on accepting a single audio clip as input. One more detail from the same documentation: 1080p in Seedance 2.5 comes out in 10-bit color and the H.265 codec, so if a file will not open, play it in VLC, mpv or QuickTime. Neither version accepts references with real people's faces directly, and for such cases BytePlus offers a separate portrait-assets mechanism (documentation).
Seedance: official site, download, subscription
Searches like "seedance download" and "seedance official site" run into confusion: the word names a model, not a product. There are three real entry points:
- The model page on seed.bytedance.com: a description and examples, no generation there.
- Dreamina (dreamina.capcut.com): ByteDance's service with a web interface. According to the Dreamina pricing page, Seedance 2.0 and 2.5 are available on a paid monthly subscription, and prices depend on the region.
- BytePlus ModelArk: the API for developers. To enable Seedance 2.0 and 2.5, the documentation requires an account balance above $30, or an AI Savings Plan from $30, or a purchased resource package.
We found no standalone Seedance app in the official materials. Sites like "seedance free online" and files like "seedance.apk" do not appear in ByteDance's materials, so never hand them a key or an account password.
Seedance 2.5 for free: what you can check
The short honest answer: we found no full free access to 2.5. Dreamina says it gives trial credits to eligible accounts, but the amounts and conditions change, so check your balance after signing in. For Seedance 2.5 specifically, it names a subscription.
To judge the quality without spending much, generate a short 4-5 second clip at 480p (BytePlus lists it among the options for 2.5) and feed in one reference instead of ten. That shows whether the model understands your scene. Turn on the full length and 1080p later. Open models that cost nothing, such as Kandinsky 6.0, are covered in the overview of AI video generators.
How to write a Seedance 2.5 prompt
The official BytePlus guide advises thinking in a visual story, not a pile of tags. First write one sentence: who, where, what happens, in what style, how it is shot. Then unfold the plot by timing. Finally set the general rules of the shot (prompt guide). Five more rules from it:
- Split time in 1-second steps: "0-3 s", "3-7 s". Gaps such as "0-3 s, 5-6 s" are not allowed.
- Give every reference a role and a number in upload order: "Image 1", "Video 1", "Audio 1".
- When editing, describe the move from A to B: what it was and what it should become.
- Phrase things as statements. Negations are supported for subtitles and sound: "no subtitles", "no music".
- If a reference is exact, do not retell it: "repeat the movements from Video 1 exactly".
The same guide also suggests how many references to take. If the characters are defined by images, 1-8 subjects work most stably: 9-12 will go through, but the result is unstable. If they are defined by audio and video, the optimum is 1-5 subjects and samples of 5-10 seconds. Edit videos of up to 20 seconds, and keep a storyboard within 15 panels.
Below are seven prompts. They are our own examples, written to the guide's rules, not vendor quotes. The guide does not name supported languages and only mentions "multilingual creation", so if you prompt in a language other than English, compare it with an English version on the first test.
Prompt 1: a product spot
"Commercial spot, soft studio light. White sneakers on wet black cobblestones at night, streetlamps reflected in the puddles. Low camera, slow push-in. 0-3 s: drops fall on the toe, lamplight slides across the mesh. 3-7 s: the camera circles the sneaker from the left, focus moves to the logo on the heel. Sound: rain and occasional footsteps far away. No music, no subtitles." It has a short setup, timing, camera and a ban on extra sound.
Prompt 2: a dialogue with references
"Image 1: Anna, voice from Audio 1. Image 2: Max, voice from Audio 2. Kitchen, morning, medium shot, light handheld camera. Anna: 'Did you forget the keys again?' Max, without turning around: 'They're in the jacket.' No subtitles." An image and a voice are tied to each character: without that, the model mixes up the characters.
Prompt 3: first and last frame
"First frame: Image 1. Last frame: Image 2. In between, a gull takes off from the pier, the camera rises after it, sunset, the sound of waves." The guide recommends passing frame roles as a request parameter: then the format is fixed by the first frame.
Prompt 4: editing a video
"In Video 1, replace the hero's red jacket with the yellow raincoat from Image 1. Movement, camera and sound unchanged." The prompt names both what changes and what stays the same.
Prompt 5: extending a clip
"Extend Video 1 forward: the hero opens the warehouse door, it is dark inside, a flashlight beam slides across the shelves. Keep the style and sound of the original clip." The guide names "extend forward" and "continue" as the extension triggers, so keep one of those phrases and test the wording on a short clip. The model takes the frame format from the source video.
Prompt 6: camera and transition
"One continuous shot, no cuts. A drone rises over a fishing village at dawn, then flies forward along the embankment. At 5 seconds, a smooth transition into an orbit around a motorboat. Only ambient sounds: gulls, wind, water." The moment and the method of the transition are stated outright, and extra sound is banned.
Prompt 7: motion transfer
"Repeat the movements and camera work from Video 1 exactly. Take the hero from Image 1, background: the courtyard from Image 2." The details of the motion are not spelled out, because the reference carries them.
Seedance 2.5 through the codexkey API
codexkey accepts video tasks on POST /v1/videos: key sk_cdx_..., address https://api.codexkey.online/v1, model seedance-2.5. The status is checked through GET /v1/videos/{id}, and the finished file is fetched from GET /v1/videos/{id}/content. The format is compatible with the OpenAI Videos API, and the details are in the images and video reference.
curl https://api.codexkey.online/v1/videos \
-H "Authorization: Bearer sk_cdx_YOUR_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "seedance-2.5", "prompt": "Commercial spot, soft light. White sneakers on wet cobblestones at night. Low camera, slow push-in. No music, no subtitles."}'
You create the key in your account, and the steps are in the quickstart. The response carries the task id: put it into the status request and, once the task finishes, download the file. Allowed duration, resolution and modes change, so take them from the model page in the catalog. The public price list has no video prices; the price is shown on the model page. Money is charged only for a successful result. Payment is in rubles: a Russian bank card, SBP (the Russian Faster Payments System) or cryptocurrency, with no VPN and no account at the vendor.
For a comparison with other video models, from Kling to Veo, see the overview of AI video generators. If you doubt the choice, run the same prompt on Seedance 2.5 and on Kling AI 3.0 and compare the results. If you need a clip with sound from Google, read how to use Veo 3.1. And what Kling's free tier actually gives you is covered in Kling AI free credits, limits and alternatives.
FAQ
Is Seedance 2.5 free: can I use it without paying?
We found no fully free access. Dreamina gives trial credits to eligible accounts, but for Seedance 2.5 the official pricing page names a paid monthly subscription. Check the size of the credits in your own account.
Where can I download Seedance: is there an app?
ByteDance's official materials show no Seedance app: it is the name of a model. Work through the Dreamina web service or the BytePlus API. Files like "seedance.apk" and "download seedance" offers on third-party sites do not appear in ByteDance's official materials.
Which Seedance site is the official one?
The model page: seed.bytedance.com. The generation service: Dreamina. The API for developers: BytePlus ModelArk. Other addresses, including "seedance free online", do not appear in ByteDance's official materials, so it is better not to hand them payment details or passwords.
How does Seedance 2.5 differ from Seedance 2.0?
Version 2.5 gives a clip up to 30 seconds instead of 15 and accepts up to 50 references instead of 15. On the other hand, its resolution ceiling is 1080p, while 4K stayed with 2.0 in the BytePlus documentation.
How do I write a prompt for Seedance 2.5?
Start with one sentence: who, where, what happens, in what style and how it is shot. Then split the plot by seconds, assign roles to the references ("Image 1", "Audio 1") and rule out the extras right away: "no subtitles", "no music". Example: "White sneakers on wet cobblestones at night. 0-3 s: drops fall on the toe. 3-7 s: the camera circles the sneaker from the left. No music, no subtitles".
Seedance API: how do I connect from Russia?
Going straight through BytePlus takes an account and one of these conditions: a balance above $30, an AI Savings Plan from $30, or a purchased resource package. Through codexkey a key sk_cdx_... is enough: requests go to POST /v1/videos with the model seedance-2.5, and you pay by Russian bank card, SBP or cryptocurrency.
How much does Seedance 2.5 cost?
Dreamina's prices depend on the region and the subscription. codexkey has no public video price list, the price is shown on the model page, and money is charged only for a successful result.