提交任务

prompt、image、duration、size 等字段全部平铺在请求体顶层，没有 content 数组，也没有 metadata 包装。图片放进 image 单字段，固定按首帧处理。其余参数（fps、seed、response_format、callback_url）同样在顶层。

与豆包原生的差异见下方「与原生结构的差异」。

prompt, image, duration, size, and the other fields all sit flat at the top level, with no content array and no metadata wrapper. An image goes in the single image field and is always treated as the first frame. The other parameters (fps, seed, response_format, callback_url) are likewise at the top level.

For the differences from Doubao native, see "Differences from the native structure" below.

OpenAI 兼容结构只有 prompt 与 image 两个创作字段，能表达的场景比豆包原生少。需要首尾帧、参考图、视频编辑或延长时，用豆包原生通道提交，任务查询仍然走同一个任务状态接口。

The OpenAI-compatible schema has only two creative fields, prompt and image, so it expresses fewer scenarios than Doubao native. For first-and-last frames, reference images, video editing, or extension, submit through the Doubao native channel; task polling still uses the same task status endpoint.

更多状态码与重试建议见错误码与限流，常见报错的排查步骤见常见问题。

For more status codes and retry guidance see Error codes and rate limits; for troubleshooting steps on common errors see the FAQ.

JSON 请求体不能直接携带二进制文件。首帧图如果要本地直传，改用 multipart/form-data：

curl --request POST   --url 'https://ark.oairoute.com/v1/video/generations'   --header 'Authorization: Bearer YOUR_ARK_KEY'   --form 'model=seedance-2.5'   --form 'prompt=海面日落，镜头缓慢前推'   --form '[email protected]'   --form 'duration=5'   --form 'size=1280x720'

两个容易踩的坑：

• 数字要写成字符串。 multipart 里没有类型概念，duration=5 与 duration="5" 传输上一样，但部分客户端会把 5 序列化成 5.0，被判为非法整数
• image 与 image_url 互斥。 同时出现返回 400

multipart 的请求体比 JSON 大，10MB 以上的图片在慢网络下容易超时，建议改用先上传再引用。

A JSON body can't carry binary files directly. To upload a first-frame image from disk, use multipart/form-data:

curl --request POST   --url 'https://ark.oairoute.com/v1/video/generations'   --header 'Authorization: Bearer YOUR_ARK_KEY'   --form 'model=seedance-2.5'   --form 'prompt=Sunset over the sea, camera slowly pushing in'   --form '[email protected]'   --form 'duration=5'   --form 'size=1280x720'

Two common traps:

• Write numbers as strings. Multipart has no type system, so duration=5 and duration="5" transmit the same; but some clients serialize 5 as 5.0, which is rejected as an invalid integer
• image and image_url are mutually exclusive. Passing both returns a 400

Multipart bodies are larger than JSON, and images over 10MB time out easily on slow networks; uploading first and referencing is recommended.

metadata 是自由格式对象，服务端不校验内部键名，只做两件事：原样保存、在任务查询时原样返回。

{
  "model": "seedance-2.5",
  "prompt": "海面日落",
  "metadata": {
    "order_id": "A20261030-001",
    "user_id": "u_88213"
  }
}

序列化后不超过 4KB，超限返回 400。

metadata 不进入模型，写进去的「让画面更亮一点」不会影响生成结果。需要影响生成的内容一律放 prompt。

metadata is a free-form object; the server doesn't validate its keys and does only two things: store it verbatim and return it verbatim on task queries.

{
  "model": "seedance-2.5",
  "prompt": "Sunset over the sea",
  "metadata": {
    "order_id": "A20261030-001",
    "user_id": "u_88213"
  }
}

It must fit in 4KB after serialization; exceeding that returns a 400.

metadata never reaches the model; writing "make it brighter" there won't affect the generation. Anything that should influence generation belongs in prompt.

以下视频模型均可通过本接口提交任务，清单只列各厂商最新一代，实际可用范围以你的账号分组为准：

各型号的能力差异：

2.0 家族里 fast 与 mini 只支持 480p 与 720p，4K 仅 2.0 标准版可用。

Every video model below can be submitted through this endpoint; the list shows only each vendor's latest generation, and what your account group actually allows takes precedence:

Differences between model families:

Within the 2.0 family, fast and mini support only 480p and 720p; 4K is available only on the 2.0 standard model.

文生视频

OpenAI 兼容结构把文本放进 prompt，图片放进 image，比原生通道少一层 content 数组包装。提交返回 task_id，再轮询任务状态接口。

图生视频

image 固定按首帧处理。需要控制末帧请改用豆包原生通道。

指定时长与帧率

duration 单位是秒，fps 决定流畅度。时长越长、帧率越高，耗时和费用都会上去。

回调通知

大批量任务不要靠轮询，传 callback_url，任务完成时平台会主动 POST 结果过来。

Text to video

The OpenAI-compatible schema puts text in prompt and an image in image, one layer of content-array wrapping less than the native channel. Submission returns a task_id; poll the task status endpoint from there.

Image to video

image is always treated as the first frame. To control the last frame, use the Doubao native channel instead.

Setting duration and frame rate

duration is in seconds and fps determines smoothness. Longer duration and higher frame rate both raise time and cost.

Callback notification

For large batches, don't rely on polling — pass callback_url and the platform POSTs the result when the task completes.