SeeDance Video Generation

SeeDance is the TENSORAXIS integration of the Volcengine Ark Doubao Seedance video models (channel doubao-video). Like every video model it is asynchronous: submit a task to get a task_id, poll the task status, then read the video URL after completion. The shared submit/poll flow lives in the Video Generation overview; this page covers only the SeeDance-specific capabilities and parameters.

Models and Capability Matrix

2-0-fast matches 2-0 in capability but is optimized for faster generation. Capability support is ultimately defined by the upstream model.

Request Fields

POST /v1/video/generations

Text-to-Video

Provide only prompt; add resolution, duration, aspect ratio, etc. via metadata.

curl https://api.tensoraxis.ai/v1/video/generations \
  -H "Authorization: Bearer $TENSORAXIS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "doubao-seedance-2-0-260128",
    "prompt": "An old man wearing a hat smiles and walks forward down an autumn street",
    "metadata": {
      "resolution": "1080p",
      "ratio": "16:9",
      "duration": 5
    }
  }'

Image-to-Video

First-frame

Pass a reference image via images to use it as the first frame; other parameters match text-to-video.

curl https://api.tensoraxis.ai/v1/video/generations \
  -H "Authorization: Bearer $TENSORAXIS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "doubao-seedance-1-0-lite-i2v",
    "prompt": "Slow push-in, the subject smiles",
    "images": ["https://example.com/first-frame.jpg"],
    "metadata": {
      "resolution": "1080p",
      "duration": 5
    }
  }'

First/last frame (Seedance 2.0)

First/last frame requires metadata.content to tag each image with an explicit role (first_frame / last_frame); it overrides the content list generated from images. See metadata.content shape.

{
  "model": "doubao-seedance-2-0-260128",
  "prompt": "A time-lapse from dawn to sunset",
  "metadata": {
    "resolution": "1080p",
    "duration": 5,
    "content": [
      { "type": "image_url", "image_url": { "url": "https://example.com/start.jpg" }, "role": "first_frame" },
      { "type": "image_url", "image_url": { "url": "https://example.com/end.jpg" }, "role": "last_frame" }
    ]
  }
}

Video Reference / Continuation (equivalent to "video edit")

The SeeDance backend has no dedicated frame-level video-edit endpoint. Seedance 2.0's "video multimodal reference" capability lets you feed a video as a reference/continuation source — pass a metadata.content item with video_url.

{
  "model": "doubao-seedance-2-0-260128",
  "prompt": "Continue the motion in this clip, the camera keeps panning right",
  "metadata": {
    "resolution": "1080p",
    "content": [
      { "type": "video_url", "video_url": { "url": "https://example.com/source.mp4" } }
    ]
  }
}

metadata Advanced Parameters

Fields under metadata map to top-level fields of the Volcengine task. The "Range / Default" column comes from the official Volcengine documentation and is ultimately defined by the upstream model; TENSORAXIS only receives and forwards these fields.

Notes:

• metadata.model is removed and cannot override the billed model.
• metadata keys not mapped to the fields above are usually not forwarded upstream.
• If both seconds and metadata.duration are sent, metadata.duration wins.

metadata.content shape

metadata.content is an advanced content array used for first/last frame, video reference, audio reference, etc. Once provided, it overrides the content list generated from images (any type: text item is ignored; the body comes from prompt).

Billing

SeeDance is billed by token; resolution is already reflected in the token count, so no separate resolution coefficient applies.

• Plain generation (text/image-to-video) uses the generate tier.
• Video input (video reference/continuation) is detected from the real media content and falls into the cheaper video_input discount tier (lower unit price than plain generation).

Exact unit prices follow the console "Video Pricing" configuration and the pricing page; operators may override the defaults with the latest official rates.

Task Polling

The shared submit/poll flow, status codes, and OpenAI/Sora-compatible fetch paths are in the Video Generation overview.