API Reference

Seedance

Seedance is an advanced video generation model series supporting text-to-video and image-to-video generation, with multi-modal reference inputs, video editing, and video extension available since the 2.0 series. SIRAYA-Seedance-2.5 extends this further with audio-only references and coherent single-shot videos up to 30 seconds long.

POST https://llm.siraya.ai/v1/videos/generations

Supported Models

Model ID	Description	Resolutions
`ByteDance-Seedance-1.0-pro`	Base version	480p, 720p, 1080p
`ByteDance-Seedance-1.0-pro-fast`	Base fast version	480p, 720p, 1080p
`ByteDance-Seedance-1.5-pro`	Advanced version	480p, 720p, 1080p
`SIRAYA-Seedance-2.0`	Supports multi-modal references	480p, 720p, 1080p, 4k
`SIRAYA-Seedance-2.0-fast`	Fast version, supports multi-modal references	480p, 720p
`SIRAYA-Seedance-2.0-mini`	Lower-cost version, supports multi-modal references	480p, 720p
`SIRAYA-Seedance-2.5`	Latest version — up to 30s videos, audio-only references, `mov` output, 50 reference assets	480p, 720p, 1080p

Authorization string Required

Your API Key (e.g., Bearer <API_KEY>).

Body

model string Required

The model ID (e.g., ByteDance-Seedance-1.5-pro, SIRAYA-Seedance-2.0, SIRAYA-Seedance-2.5).

prompt string Required

A text description for the video content.

seconds integer

The duration of the video in seconds. Supported values: 4 to 15 (4 to 30 for SIRAYA-Seedance-2.5), or -1 to let the model choose automatically. When input_references includes a video reference, pass -1 so the output length follows the input video — see Video Editing & Extension. duration is not a supported field; use seconds. An unsupported value is accepted at submit and fails later, when the job runs — validate it client-side (see Failures after submit).

resolution string

480p, 720p, 1080p, or 4k. The values each model accepts are listed under Supported Models. An unsupported value is not rejected at submit; the task fails afterwards (see Failures after submit).

aspect_ratio string

Aspect ratio of the video: 16:9, 4:3, 1:1, 3:4, 9:16, 21:9, or adaptive (model picks automatically). For video editing, video extension, and first-frame/first-last-frame requests, the Seedance 2.0 family accepts a specified ratio or adaptive, while SIRAYA-Seedance-2.5 only supports adaptive — see Video Editing & Extension.

image_url string

Deprecated — prefer frame_images. Input image URL for image-to-video generation; it is treated as a single first_frame, so it cannot be combined with input_references.

seed integer

Random seed for reproducibility. The same seed produces the same video.

generate_audio boolean

Whether to generate an audio track for the video. SIRAYA-Seedance-2.5 natively supports prompts and generated audio in 11 languages: Chinese, English, Spanish, Indonesian, Malay, Thai, Arabic, Portuguese, Vietnamese, Japanese, and Korean.

frame_images array

First-frame and/or last-frame image control. Each item has a frame_type (first_frame or last_frame) and an image_url. Cannot be used together with input_references.

input_references array

Multi-modal reference materials (image, video, audio) to influence style or content, edit an existing video, or extend one. Supported by SIRAYA-Seedance-2.0, SIRAYA-Seedance-2.0-fast, SIRAYA-Seedance-2.0-mini, and SIRAYA-Seedance-2.5. Cannot be used together with frame_images. See input_references for per-model limits.

output_format string Default: mp4

Output video container: mp4 (standard, broadest playback compatibility) or mov (higher color precision — recommended for editing/extension workflows). SIRAYA-Seedance-2.5 only.

extra_body map

Specific parameters.

async boolean Default: false

When false, the request waits for generation to finish and returns the completed video URL. When true, it returns immediately with a video id you poll for status. See Asynchronous generation.

callbackUrl string

Only used when async is true. An http(s) URL the router will POST to once generation completes (or fails), so you don't have to poll.

cURLPython

curl https://llm.siraya.ai/v1/videos/generations \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <API_KEY>" \
  -d '{
    "model": "ByteDance-Seedance-1.5-pro",
    "prompt": "A shiba inu running on the beach",
    "resolution": "720p",
    "aspect_ratio": "16:9",
    "seconds": 5
  }'

import requests

url = "https://llm.siraya.ai/v1/videos/generations"
headers = {
    "Content-Type": "application/json",
    "Authorization": "Bearer <API_KEY>"
}
data = {
    "model": "ByteDance-Seedance-1.5-pro",
    "prompt": "A shiba inu running on the beach",
    "resolution": "720p",
    "aspect_ratio": "16:9",
    "seconds": 5
}

response = requests.post(url, headers=headers, json=data)
print(response.json())

Example Response

{
    "created": 1760347750,
    "data": [
        {
            "url": "https://resources.siraya.ai/v1/videos/video_n6Sn-1aIT3z8x.../content"
        }
    ],
    "error": {}
}

Mode Exclusivity

The following three modes are mutually exclusive — only one can be used per request:

Mode	Parameter	Use Case
First-frame control	`frame_images` (`first_frame`), or the deprecated `image_url`	Pin the exact first frame
First + last frame control	`frame_images` (`first_frame` + `last_frame`)	Pin both start and end frames
Multi-modal reference	`input_references`	Style/content reference, video editing, or video extension

Note: If you need both reference materials and first-frame control, use input_references and describe the desired first frame in your prompt. The model will follow the description but cannot guarantee pixel-level accuracy. Use frame_images when exact frame consistency is required.

frame_images

Specifies the starting and/or ending frame images for the video. The AI generates content between the provided frames.

"frame_images": [
  {
    "frame_type": "first_frame",
    "image_url": "https://your-domain.com/start.jpg"
  }
]

Field	Description
`frame_type`	`first_frame` (start frame) or `last_frame` (end frame)
`image_url`	Publicly accessible URL of the image

input_references (Seedance 2.0 and 2.5)

Pass image, video, or audio files as generation references via input_references. Supported by SIRAYA-Seedance-2.0, SIRAYA-Seedance-2.0-fast, SIRAYA-Seedance-2.0-mini, and SIRAYA-Seedance-2.5.

Reference material with a human face

If the reference image or video contains a human face (not a celebrity or public figure), you can't pass a public URL — it must first be uploaded as a trusted asset and referenced as asset://{assetId}. See Using Assets with Seedance 2.0 for the full workflow, including motion-reference and character-consistency use cases.

"input_references": [
  {
    "type": "image",
    "url": "https://your-domain.com/character.jpg",
    "role": "reference_image"
  }
]

Field	Description
`type`	`image`, `video`, or `audio`
`url`	Publicly accessible URL of the reference material
`role`	`reference_image`, `reference_video`, or `reference_audio`

Reference limits by model:

Model	Max references	Audio-only references
`SIRAYA-Seedance-2.0` / `SIRAYA-Seedance-2.0-fast` / `SIRAYA-Seedance-2.0-mini`	15 total (9 images + 3 videos + 3 audio)	Not supported — needs at least one image or video
`SIRAYA-Seedance-2.5`	50 total (30 images + 10 videos + 10 audio)	Supported

Video Editing & Extension (Seedance 2.0 and 2.5)

SIRAYA-Seedance-2.0 (and -fast/-mini) and SIRAYA-Seedance-2.5 classify every input_references request into one of five task types, based on the reference roles and the intent expressed in prompt: text-to-video, reference-to-video, video editing, video extension, or first-frame/first-last-frame generation. Editing and extension are triggered the same way on both families — pass a reference_image/reference_video/reference_audio and phrase the prompt accordingly. Where the two families differ is aspect_ratio flexibility:

Task	Trigger	`aspect_ratio` — Seedance 2.0 family	`aspect_ratio` — `SIRAYA-Seedance-2.5`
Video editing	A reference + a prompt with an editing verb (e.g. "edit", "add", "remove", "replace")	A specified ratio, or `adaptive` to match the input video	Only `adaptive` — matches the input video
Video extension	A reference + a prompt with an extension verb (e.g. "extend", "continue")	A specified ratio, or `adaptive` to match the input video	Only `adaptive` — matches the input video
First-frame(s) generation	`frame_images` with `first_frame` and/or `last_frame`	A specified ratio, or `adaptive` to match the first-frame image	Only `adaptive` — matches the first frame
Text-to-video / reference-to-video	No editing or extension wording	Any supported value	Any supported value

Recommended: pass seconds: -1 when a video reference is present

If input_references includes any item with "type": "video" — video editing, video extension, or any reference-to-video request with a video reference — set seconds to -1. The output duration is then determined automatically from the input video (matched for editing, extended for extension). If you omit seconds, the model's default length is used instead, which is not matched to the input video, and a fixed value may be rejected for the classified task type (see below).

Model group	Request includes a video reference	No video reference (text-to-video, image/audio-only)
Seedance 2.0 family (`SIRAYA-Seedance-2.0` / `-fast` / `-mini`)	`-1` (recommended)	`4`–`15`, or omit to let the model choose
`SIRAYA-Seedance-2.5`	`-1` (recommended)	`4`–`30`, or omit to let the model choose

Failures after submit

Seedance validates much of the request only after the job starts, so a 200 at submit does not mean the render will succeed. The job then ends with status: failed (or, in synchronous mode, a 400 with the error message — see Asynchronous generation). Handle failed when polling or in your callback. Typical causes:

Misclassification. An unsupported aspect_ratio/seconds combination for the classified task type doesn't fail immediately — the request queues normally, and the error is returned once the model finishes classifying the task. For reference-to-video requests, avoid editing/extension wording in your prompt, or the model may misclassify the task and reject it.
Unsupported seconds or resolution values. These are not checked at submit.
Content checks on the output. A benign prompt can still fail when the generated audio or video is flagged (for example with an OutputAudioSensitiveContentDetected policy error), including on 4k renders of SIRAYA-Seedance-2.0. Retry with a different prompt or settings.

{
  "model": "SIRAYA-Seedance-2.5",
  "prompt": "Video edit: remove everyone except the main subject.",
  "input_references": [
    {
      "type": "video",
      "url": "https://your-domain.com/original.mp4",
      "role": "reference_video"
    }
  ],
  "aspect_ratio": "adaptive",
  "seconds": -1,
  "output_format": "mov"
}