Seedance SG Series

The SG series provides four open models to support MoonNexAI generic video parameters compatible with Volcengine-compatible. Both access points use MoonNexAI API Key, model always fill out the public model name in the following table.

Models and capabilities

Open Model	resolution	Integer duration	Picture / Video / Audio Configuration Cap
seedance-2.0-sg-mini	480P, 720P	4 - 15 seconds	9 / 3 / 3 Total 15
seedance-2.0-sg	480P, 720P, 1080P, 4K	4 - 15 seconds	9 / 3 / 3 Total 15
seedance-2.0-sg-fast	480P, 720P	4 - 15 seconds	9 / 3 / 3 Total 15
seedance-2.5-sg	480P, 720P, 1080P	4 - 30 seconds	30 / 10 / 10 Total 50

Configure allows duration=-1 to automatically select duration; pass the integer number of seconds when target duration is to be determined. aspect ratio supports 16:9, 9:16, 1:1, 4:3, 3:4, 21:9, adaptive depending on the mode of generation. Current account available models and capabilities are based on models API and console.

All four models have been verified at 720P, 4 seconds with one reference_image. Earlier validation also covered 2 images + 1 audio file + 1 video at resolutions available then. Standard 1080P and 4K two-image results are listed below. This does not establish that every count limit, automatic duration or other duration has been individually tested.

seedance-2.0-sg has also been verified with compatible content[] containing 6 images + 1 MP3 audio file and no video reference. A 480p, 16:9, 10 second request with generate_audio=true succeeded and returned an MP4 with audio, passing download and full decoding checks. This verifies only that combination, not every model, maximum reference count or audio format.

Standard 1080P / 4K

On 2026-09-21, seedance-2.0-sg completed two real generations through POST /v1/videos. Both used 2 photos of real people, role=reference_image, image_source_mode=asset, aspect_ratio=16:9, duration=4 and generate_audio=false, without video or audio references.

Requested resolution	Successful response resolution	Observed MP4 dimensions	Observed file duration
1080P	1080p	1920×1080	About 4.04 seconds
4K	4k	3840×2160	About 4.04 seconds

Both outputs passed download and full decoding checks. The dimensions and durations shown are observations from these tasks, not fixed output guarantees. They do not establish the same support for Fast, Mini or other reference combinations.

Reference Mode and media assets

Use role=reference_image for images, reference_video for videos and reference_audio for audio. Keep reference_image even for a single reference image; do not change it to first_frame. Prompt markers such as @图片1 describe references but do not replace the structured role field.

• Audio requires an accompanying image or video. If an image is already present, adding a video just to enable audio is unnecessary.
• First/last frames use a separate mode that cannot be mixed with multimodal references. A last frame requires a first frame. Use reference_image for reference requests requiring a fixed output aspect ratio.
• media assets URL uses an anonymous downloadable HTTPS media straight chain, which continues to be valid until task is completed; it cannot be a login page or preview page. Pictures, audio and video still need to meet content security and media specifications.
• The web image URL can be submitted directly, with media assets preparation required for service completion, without the need for a separate authentication number.
• MoonNexAI Asset:// references must be current account and suitable for target models. Authentication bindings to be prepared, failed or not matched should be processed first; do not repeat the authentication numbers of other models. The following validated processes use the URL of the public network, for which the independent creation of authentication media assets is described in Uploads and assets.

Picture media assets Mode

When SG request contains pictures and image_source_mode is not transmitted, MoonNexAI uses asset by default. This mode prepares pictures media assets for subsequent submission and recommends for real person photographs or other content that could trigger a photo privacy audit.

image_source_mode	Behaviour	Apply scene
asset	Submitted by service preparation of photo media assets.	default value;reals, characters or images that require more stable media assets processing.
direct_url	Use the URL of the public web image in request directly.	The photo can be rejected by a private audit.

image_source_mode applies to all images in a request; do not mix modes across images in one task. Avatar assets prepared through the SG workflow can be supplied as avatar_asset_id or avatar_asset_ids. These fields also imply asset when image_source_mode is omitted.

MoonNexAI common parameters

The recommended entry is POST /v1/videos. The following is an example of 2 + 1 audio + 1 video, 1080P and 4 seconds; the example.com address is a placeholder and needs to be replaced by the real media assets straight chain you are entitled to use.

curl https://moonnexai.com/v1/videos \
  -H "Authorization: Bearer <MOONNEXAI_API_KEY>" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "seedance-2.0-sg",
    "prompt": "保持 @图片1 与 @图片2 的主体外观，参考 @视频1 的镜头运动，结合 @音频1 的声音氛围。",
    "image_source_mode": "asset",
    "references": [
      { "media_type": "image", "role": "reference_image", "url": "https://example.com/reference-1.png" },
      { "media_type": "image", "role": "reference_image", "url": "https://example.com/reference-2.png" },
      { "media_type": "audio", "role": "reference_audio", "url": "https://example.com/reference-audio.wav" },
      { "media_type": "video", "role": "reference_video", "url": "https://example.com/reference-video.mp4" }
    ],
    "size": "1080p",
    "aspect_ratio": "16:9",
    "duration": 4,
    "generate_audio": true
  }'

For single graph references, only the first picture entry is retained, and prompt is changed to "Reference @pho1, keep the main look and the lens is smoothly moving." And role=reference_image. When SG 2.0 Fast or Mini is selected, model is changed to the corresponding public name and size is changed to 480p or 720p; the standard version can also use 1080p or 4K. 480p, 720p or 1080p is used for selecting 2.5.

Official compatibility parameters

The compatible endpoint is POST /api/v3/contents/generations/tasks; POST /v1/videos accepts the same compatible body. Choose one format per request. Avoid sending both references[] and content[], or duplicating the prompt.

MoonNexAI Universal field	Officially field	Annotations
model	model	Fill in SG public model names
prompt	type=text Entry for content[]	Text prompt
references[]	Media entry for content[]	Retain the reference role in a prominent way
size	resolution	Select resolution by Model
aspect_ratio	ratio	Output aspect ratio
duration	duration	Integer seconds, or -1 when supported
generate_audio	generate_audio	boolean values;invisible false off-generation tracks

The following is the same media assets and target specification as the generic parameter examples:

curl https://moonnexai.com/api/v3/contents/generations/tasks \
  -H "Authorization: Bearer <MOONNEXAI_API_KEY>" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "seedance-2.5-sg",
    "content": [
      { "type": "text", "text": "保持 @图片1 与 @图片2 的主体外观，参考 @视频1 的镜头运动，结合 @音频1 的声音氛围。" },
      { "type": "image_url", "role": "reference_image", "image_url": { "url": "https://example.com/reference-1.png" } },
      { "type": "image_url", "role": "reference_image", "image_url": { "url": "https://example.com/reference-2.png" } },
      { "type": "audio_url", "role": "reference_audio", "audio_url": { "url": "https://example.com/reference-audio.wav" } },
      { "type": "video_url", "role": "reference_video", "video_url": { "url": "https://example.com/reference-video.mp4" } }
    ],
    "resolution": "1080p",
    "ratio": "16:9",
    "duration": 4,
    "generate_audio": true
  }'

Compatibility covers request fields and paths. It does not mean all advanced parameters, editing or extension modes are verified. Follow the examples above for standard reference generation.

6 pictures + 1 audio, no reference video

The example preserves the verified media counts, roles and target specifications, with placeholder prompts and URLs. Replace all 7 example.com addresses with real, anonymously downloadable media that you may use. Images map to @图片1 through @图片6 by their image order in content[]; audio maps to @音频1.

curl https://moonnexai.com/api/v3/contents/generations/tasks \
  -H "Authorization: Bearer <MOONNEXAI_API_KEY>" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "seedance-2.0-sg",
    "content": [
      { "type": "text", "text": "以 @图片1 与 @图片6 为角色参考，保留 @图片2、@图片3、@图片5 的物体特征，参考 @图片4 的环境与光照。让 @图片6 的角色使用 @音频1 的声音特征说话，镜头平稳自然。" },
      { "type": "image_url", "role": "reference_image", "image_url": { "url": "https://example.com/reference-1.png" } },
      { "type": "image_url", "role": "reference_image", "image_url": { "url": "https://example.com/reference-2.png" } },
      { "type": "image_url", "role": "reference_image", "image_url": { "url": "https://example.com/reference-3.png" } },
      { "type": "image_url", "role": "reference_image", "image_url": { "url": "https://example.com/reference-4.png" } },
      { "type": "image_url", "role": "reference_image", "image_url": { "url": "https://example.com/reference-5.png" } },
      { "type": "image_url", "role": "reference_image", "image_url": { "url": "https://example.com/reference-6.png" } },
      { "type": "audio_url", "role": "reference_audio", "audio_url": { "url": "https://example.com/reference-audio.mp3" } }
    ],
    "duration": 10,
    "execution_expires_after": 3600,
    "generate_audio": true,
    "ratio": "16:9",
    "resolution": "480p",
    "watermark": false
  }'

The sound references in prompt are used to describe the intent to create it, and cannot be used as a guarantee of full consistency in voice or lines. generate_audio=true means request to generate a track; the actual sound content is based on the composition.

request and finished product specifications

resolution, ratio and duration specify targets; response fields do not replace inspecting the resulting file. The 480p / 16:9 / 10 second request above produced an observed 864×496 / 10.08 second MP4 with H.264 video and AAC audio. These are observed dimensions, not fixed guarantees. For exact dimensions, aspect ratio or duration, inspect the downloaded media and process it as your application requires.

Query and Results URL

Create response means task only. Save Open id or task_id with the same API Key polling:

curl https://moonnexai.com/v1/videos/{task_id} \
  -H "Authorization: Bearer <MOONNEXAI_API_KEY>"

GET /api/v3/contents/generations/tasks/{task_id} is used for official compatible queries. Check status and error, line up or continue polling in line, and process errors when a failure is made. video_url, url or response-compatible content.video_url are only read if the result is final and no error; the success cannot be judged by the existence of URLs alone. Details of the status and response format are provided in Query Video Job.

The URL of the downloadable media that is possible in the successful result returns as it is and can be used by the client. Please keep the address and query parameters in their entirety, without spelling out domain names and without deleting signatures; the address may be valid and should be downloaded or saved in time. If response returns to the MoonNexAI content downloading portal, then the MoonNexAI API Key is carried as required by API; do not send this Key when accessing the external media direct link.

SG GET /v1/videos/{task_id} uses queued, in_progress, completed or failed. The following shows selected fields from a successful Standard 4K response; task IDs and media URLs are placeholders:

{
  "id": "task_xxx",
  "object": "video",
  "model": "seedance-2.0-sg",
  "status": "completed",
  "resolution": "4k",
  "duration": 4,
  "url": "https://example.com/result.mp4"
}

Successful Standard 1080P tasks return resolution: "1080p"; 4K tasks return "4k". Resolution labels are normalized to lowercase and appear only when valid resolution data is available. They do not prove success. Response duration is in seconds; inspect downloaded media for exact dimensions or file duration.

billing and error

The SG series is valued at millions of output Token, and is ultimately settled at usage.completion_tokens. billing slotting model, resolution, and whether or not to enter reference video; generate_audio switches do not equal "to enter reference video." This series is based on the official original caliber, and account actually applies a unit price to read Console and price interface and documents do not have fixed amounts.

Submission may preauthorize quota, followed by an additional charge or refund based on actual usage. An asynchronous-task refund may be an adjustment on a successful task. Reconcile task status, consumption and refund records to determine the net charge. A submission failure that creates no task is not charged; failures after creation follow the complete billing record.

For parameter or asset failures, inspect the safe error and available code and param fields, and retain the MoonNexAI request ID. If a reference image is treated as a first frame, check its explicit reference_image role and whether first/last-frame roles were mixed in. Distinguish submission failure from failure after task creation to avoid unnecessary duplicate charges.

Relevant Pages

• Seedance Video Overview
• Video Model Matrix
• Create Video Job
• Query Video Job