Seedance 2.5 video generation API
seedance2.5 uses the same async task contract and supports higher reference limits for images, video, and audio than 2.0.
Generation modes
The model determines text, first/last frame, or reference mode from references. There is no separate mode field.
| Mode | references rule |
|---|---|
| Text to video | Empty or omitted |
| First and last frame | Exactly two images with first_frame and last_frame roles |
| Reference to video | One or more role=reference image, video, or audio items |
Reference limits
| Material | Limit |
|---|---|
| Reference images | 30 |
| Reference videos | 10 |
| Reference audio | 10 |
{
"model": "seedance2.5",
"prompt": "The girl in @图片1 turns to the camera; voice follows @音频1 and motion follows @视频1",
"seconds": 5,
"resolution": "720p",
"aspect_ratio": "16:9",
"references": [
{
"type": "image",
"role": "reference",
"source": "https://cdn.example.com/girl.png"
},
{
"type": "audio",
"role": "reference",
"source": "https://cdn.example.com/voice.mp3"
},
{
"type": "video",
"role": "reference",
"source": "https://cdn.example.com/motion.mp4"
}
]
}The public model ID stays stable
The model field in the task response is the public model ID from the request. Clients do not depend on an upstream internal model name.
Reference material rules
Review HTTPS URLs, material order, and prompt references.
Task error handling
Handle polling, idempotency, and 429.