Seedance 2.5 - AnyFast

Seedance 2.5 is a multimodal video generation model for coherent long-form shots, multimodal references, precise editing, video extension, and native multilingual audio. Use the AnyFast model ID seedance-2.5 with POST /v1/video/generations.

For specific content requests, select the seedance-2.5-nsfw model under the Direct resource group. The NSFW model ID adds the -nsfw suffix to the standard model ID.

What’s new

Generate one coherent video up to 30 seconds long.
Reference up to 50 media items: 30 images, 10 videos, and 10 audio clips.
Generate from audio references alone, without an image or video reference.
Let the model choose aspect ratio and duration with ratio: "adaptive" and duration: -1.
Preserve source dimensions and timing automatically in editing, extension, and frame-guided workflows.
Generate 1080p output with 10-bit color depth and H.265/HEVC encoding.
Generate spoken or sung audio in 11 supported languages.

Capability	Seedance 2.5 limit
Output duration	`-1` for automatic duration, or an integer from 4–30 seconds
Reference images	Up to 30
Reference videos	Up to 10; 2–30 seconds each, 30 seconds total
Reference audio	Up to 10; 2–30 seconds each, 30 seconds total
Audio-only reference	Supported
Resolution	480p, 720p, 1080p, or 4K; 1080p uses 10-bit H.265/HEVC
Aspect ratio	21:9, 16:9, 4:3, 1:1, 3:4, 9:16, or adaptive
Synchronized audio	Supported
Web search	Supported for text-to-video
Return final frame	Supported
Draft-to-final workflow	Supported: create a 480p Draft, then render a 1080p Final from that Draft

Capability examples

A time-coded prompt can divide one continuous story into several acts. This example moves from a brass clock and mechanical city to an ocean voyage and a moonlit ridge, then returns to the opening clock face.

Reference image for the 30-second steampunk narrative — Reference image

30-second output

This advertising example uses one product image and six motion/composition videos. In the prompt, @image1 defines the product while @video1 through @video6 assign composition, camera movement, impact, rhythm, typography, and final framing.

Product reference image for the multimodal advertising example — Product reference image

View the six reference videos

Reference video 1 — composition

Reference video 2 — movement

Reference video 3 — impact

Reference video 4 — rhythm

Reference video 5 — typography

Reference video 6 — final framing

Combined multimodal output

The prompt removes background people while preserving the main subject. With ratio: "adaptive" and duration: -1, the result keeps the input video’s 5:4 ratio and approximately 16-second duration.

Input video (MOV)

Edited output (MOV)

Task types and classification

Seedance 2.5 classifies each request from the media role values and the prompt intent. You can also use omni_reference_task_type to select the reference, editing, or extension workflow explicitly. It supports five task families.

Task	How it is detected	Aspect ratio	Duration
Text-to-video	Text prompt without reference media	Explicit ratio or `adaptive`	`-1` or 4–30
Reference-to-video	One or more `reference_image`, `reference_video`, or `reference_audio` items	Explicit ratio or `adaptive`	`-1` or 4–30
Video editing	Reference media plus editing intent such as add, remove, replace, modify, or repair	Must be `adaptive`	Must be `-1`
Video extension	Reference media plus extension intent such as extend forward/backward, continue, or carry on	Must be `adaptive`	`-1` or 4–30
First/last-frame generation	An image with `first_frame`, optionally followed by `last_frame`	Must be `adaptive`	`-1` or 4–30

Editing, extension, and frame-guided restrictions are checked after the model classifies the task. The create request can initially return a task ID and then fail asynchronously with InvalidParameter.TaskTypeConstraint. If several parameters are invalid, the asynchronous error may list them together.

For ordinary reference generation, avoid editing or extension phrases unless that is the intended workflow. Such phrases can cause the model to classify the request as editing or extension and apply the stricter parameter rules.

Select the reference task type explicitly

Use omni_reference_task_type to control how a request containing reference media is processed. The supported values are:

Value	Behavior	Requirements
`auto`	Infer reference, editing, or extension behavior from the media and prompt	Use when you want the model to classify the workflow
`reference`	Process the request as multimodal reference generation	Provide `reference_image`, `reference_video`, or `reference_audio`
`edit`	Process the request as video editing	Provide `reference_video` and use `ratio: "adaptive"` with `duration: -1`
`extend`	Process the request as video extension	Provide `reference_video` and use `ratio: "adaptive"`; an explicit duration must be longer than the source video

If you omit this field, the model still classifies the request from the media roles and prompt intent. Set it explicitly when you need a predictable workflow.

Smarter duration and aspect ratio control

Set ratio to adaptive and duration to -1 to let Seedance 2.5 choose output dimensions and timing. Editing, extension, and first-frame/first-and-last-frame tasks have special behavior.

Video editing

For editing, ratio defaults to and only accepts adaptive. The output follows the aspect ratio of the video selected for editing. duration defaults to and only accepts -1, so the result stays close to the source duration. The output can have a fractional duration and may be up to 0.4 seconds shorter than the source because of input-frame processing.

Video extension

For extension, ratio defaults to and only accepts adaptive. The output follows the aspect ratio of the video being extended. Set duration to -1 for automatic timing or choose an integer from 4–30.

First-frame and first-and-last-frame generation

For frame-guided generation, ratio defaults to and only accepts adaptive. The output follows the first frame’s aspect ratio. Set duration to -1 or an integer from 4–30.

Text and reference generation

For text-to-video and ordinary reference-to-video, adaptive selects the best ratio from 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16. With duration: -1, the model selects an integer duration from 4–30 seconds.

Create a task

Text to video
Multimodal reference
Edit video
Extend video
First and last frame

cURL

curl --request POST \
  --url https://www.anyfast.ai/v1/video/generations \
  --header 'Authorization: Bearer YOUR_API_KEY' \
  --header 'Content-Type: application/json' \
  --data '{
    "model": "seedance-2.5",
    "content": [{
      "type": "text",
      "text": "A 20-second cinematic tracking shot through a miniature steampunk city at golden hour"
    }],
    "generate_audio": true,
    "resolution": "1080p",
    "ratio": "16:9",
    "duration": 20,
    "output_format": "mp4"
  }'

Use @image1, @video1, and @audio1 to reference media by type and order. The text prompt is optional for reference generation.

cURL

curl --request POST \
  --url https://www.anyfast.ai/v1/video/generations \
  --header 'Authorization: Bearer YOUR_API_KEY' \
  --header 'Content-Type: application/json' \
  --data '{
    "model": "seedance-2.5",
    "omni_reference_task_type": "reference",
    "content": [
      {"type":"text","text":"@image1 performs the motion from @video1, paced to @audio1"},
      {"type":"image_url","image_url":{"url":"https://example.com/character.jpg"},"role":"reference_image"},
      {"type":"video_url","video_url":{"url":"https://example.com/motion.mp4"},"role":"reference_video"},
      {"type":"audio_url","audio_url":{"url":"https://example.com/music.mp3"},"role":"reference_audio"}
    ],
    "generate_audio": true,
    "ratio": "16:9",
    "duration": 15,
    "output_format": "mp4"
  }'

See the Multimodal references capability example above for all six inputs and the generated output.

Editing requires ratio: "adaptive" and duration: -1 so the result follows the source video.

cURL

curl --request POST \
  --url https://www.anyfast.ai/v1/video/generations \
  --header 'Authorization: Bearer YOUR_API_KEY' \
  --header 'Content-Type: application/json' \
  --data '{
    "model": "seedance-2.5",
    "omni_reference_task_type": "edit",
    "content": [
      {"type":"text","text":"Edit @video1: remove every background pedestrian while preserving the main subject"},
      {"type":"video_url","video_url":{"url":"https://example.com/source.mov"},"role":"reference_video"}
    ],
    "generate_audio": true,
    "ratio": "adaptive",
    "duration": -1,
    "output_format": "mov"
  }'

See the Video editing capability example above for the input/output comparison.

Extension requires ratio: "adaptive". Set duration to -1 or an integer from 4–30; an explicit duration must be longer than the source video.

cURL

curl --request POST \
  --url https://www.anyfast.ai/v1/video/generations \
  --header 'Authorization: Bearer YOUR_API_KEY' \
  --header 'Content-Type: application/json' \
  --data '{
    "model": "seedance-2.5",
    "omni_reference_task_type": "extend",
    "content": [
      {"type":"text","text":"Extend @video1: the camera passes through the open window into a quiet gallery"},
      {"type":"video_url","video_url":{"url":"https://example.com/source.mp4"},"role":"reference_video"}
    ],
    "generate_audio": true,
    "ratio": "adaptive",
    "duration": 11,
    "output_format": "mov"
  }'

See the Video extension capability example above for all three inputs and the output.

Frame-guided generation requires ratio: "adaptive"; the output follows the first frame’s aspect ratio.

cURL

curl --request POST \
  --url https://www.anyfast.ai/v1/video/generations \
  --header 'Authorization: Bearer YOUR_API_KEY' \
  --header 'Content-Type: application/json' \
  --data '{
    "model": "seedance-2.5",
    "content": [
      {"type":"text","text":"The subject smiles at the camera while it makes a smooth 360-degree orbit"},
      {"type":"image_url","image_url":{"url":"https://example.com/first.jpg"},"role":"first_frame"},
      {"type":"image_url","image_url":{"url":"https://example.com/last.jpg"},"role":"last_frame"}
    ],
    "generate_audio": true,
    "ratio": "adaptive",
    "duration": 5,
    "output_format": "mp4"
  }'

See the First and last frames capability example above for both frame inputs and the output.

Draft to final

Draft mode lets you review a low-resolution sample before starting the final render. It is a two-task workflow: the first request creates a 480p Draft, and the second request uses that Draft to create the 1080p Final. The two tasks are billed separately.

Create a 480p Draft

Set draft to true and resolution to 480p. Submit the prompt and all source media in this request.

cURL

curl --request POST \
  --url https://www.anyfast.ai/v1/video/generations \
  --header 'Authorization: Bearer YOUR_API_KEY' \
  --header 'Content-Type: application/json' \
  --data '{
    "model": "seedance-2.5",
    "content": [{
      "type": "text",
      "text": "A small red paper boat glides across a rain puddle at dusk"
    }],
    "draft": true,
    "resolution": "480p",
    "duration": 5,
    "ratio": "16:9",
    "generate_audio": true,
    "seed": 20,
    "output_format": "mov"
  }'

Poll the returned asyntask_... ID until data.status is SUCCESS. A successful Draft result contains data.data.draft: true, resolution: "480p", and the preview URL in data.data.content.video_url.

Create the 1080p Final

Pass the public AnyFast Draft task ID in a draft_task content item. Keep model unchanged, set draft to false, and use resolution: "1080p".

cURL

curl --request POST \
  --url https://www.anyfast.ai/v1/video/generations \
  --header 'Authorization: Bearer YOUR_API_KEY' \
  --header 'Content-Type: application/json' \
  --data '{
    "model": "seedance-2.5",
    "content": [{
      "type": "draft_task",
      "draft_task": {
        "id": "asyntask_YOUR_DRAFT_TASK_ID"
      }
    }],
    "draft": false,
    "resolution": "1080p",
    "output_format": "mov",
    "return_last_frame": true,
    "priority": 2
  }'

The Final inherits the Draft prompt, input media, duration, ratio, seed, audio setting, and reference-task type. Do not repeat those inherited fields in the Final request.

The Final request can set return_last_frame, output_format, watermark, service_tier, execution_expires_after, priority, callback_url, and safety_identifier. A Draft is valid for seven days from its creation time, and the model ID used for the Final must match the Draft.

Store the public asyntask_... Draft ID returned by AnyFast and use it for the Final request. In the queried Final result, data.data.draft_task_id can be the provider-side task ID rather than the public AnyFast ID. Fields such as execution_expires_after and safety_identifier can also be normalized by the service, so do not use exact response echoing as an idempotency check.

Draft workflow FAQ

Can I generate a 720p Draft?

No. A Draft request uses 480p; the Final render uses 1080p.

Can I change the prompt or source media in the Final request?

No. The Final inherits them from the Draft. Create a new Draft when you need to change inherited inputs or generation settings.

Does a successful Draft count as the Final video?

No. The Draft is a separate 480p preview task. Submit the second request and wait for its own SUCCESS result to obtain the 1080p Final.

Where is the returned final frame?

When return_last_frame is true, read data.data.content.last_frame_url from the successful Final task result.

Additional workflows

Audio-only reference
Web search
Watermark and final frame

Seedance 2.5 can generate a video from an audio reference without any image or video input. The text prompt is optional; when included, use @audio1 to describe how the audio should guide the scene.

cURL

curl --request POST \
  --url https://www.anyfast.ai/v1/video/generations \
  --header 'Authorization: Bearer YOUR_API_KEY' \
  --header 'Content-Type: application/json' \
  --data '{
    "model": "seedance-2.5",
    "content": [
      {"type":"text","text":"Create a night-market performance paced to @audio1"},
      {"type":"audio_url","audio_url":{"url":"https://example.com/performance.mp3"},"role":"reference_audio"}
    ],
    "generate_audio": true,
    "ratio": "16:9",
    "duration": -1,
    "output_format": "mp4"
  }'

Text-to-video tasks can let the model search for current information such as products, events, or weather before generation.

cURL

curl --request POST \
  --url https://www.anyfast.ai/v1/video/generations \
  --header 'Authorization: Bearer YOUR_API_KEY' \
  --header 'Content-Type: application/json' \
  --data '{
    "model": "seedance-2.5",
    "content": [{"type":"text","text":"Create a short travel video using today’s weather in Tokyo"}],
    "tools": [{"type":"web_search"}],
    "generate_audio": true,
    "ratio": "16:9",
    "duration": 8,
    "output_format": "mp4"
  }'

Use watermark to add the AI-generated mark and return_last_frame to request the final frame alongside the video result.

{
  "watermark": true,
  "return_last_frame": true
}

Advanced request controls

Field	Behavior
`execution_expires_after`	Maximum queued/running lifetime in seconds; default `172800`.
`priority`	Task execution priority; default `0`.
`safety_identifier`	Stable end-user identifier for safety tracing.
`seed`	Random seed; `-1` chooses a random value.
`service_tier`	Task service tier; default `default`.
`output_format`	Output video container: `mp4` (default) or `mov`. Use the exact field name `output_format`; `out_format` is not a supported alias.

A task becomes expired when it remains queued or running beyond execution_expires_after.

Output specifications

Resolution and pixel dimensions

resolution defaults to 720p. Seedance 2.5 supports 480p, 720p, 1080p, and 4k. The 1080p option uses 10-bit color depth and H.265/HEVC encoding. Some playback environments do not support H.265/HEVC, so use VLC, MPV, or QuickTime Player if the video does not play normally.

Resolution	Ratio	Pixel dimensions
480p	16:9	854×480
480p	4:3	752×560
480p	1:1	640×640
480p	3:4	560×752
480p	9:16	480×854
480p	21:9	992×432
720p	16:9	1280×720
720p	4:3	1112×834
720p	1:1	960×960
720p	3:4	834×1112
720p	9:16	720×1280
720p	21:9	1470×630
1080p	16:9	1920×1080
1080p	4:3	1664×1248
1080p	1:1	1440×1440
1080p	3:4	1248×1664
1080p	9:16	1080×1920
1080p	21:9	2206×946

Output container

Set output_format to mp4 (default) or mov. A successful task query echoes the selected value in data.data.output_format. Use the exact field name output_format; out_format is not a supported alias and may be ignored. MOV playback support varies. VLC, mpv, and ffplay support it on macOS and Windows; IINA supports it on macOS. Browser playback may not work, so download the file and use a compatible player when necessary.

Audio, watermark, and final frame

generate_audio defaults to true and enables synchronized dialogue, sound effects, and music.
watermark defaults to false. Set it to true to add an AI-generated mark in the lower-right corner.
return_last_frame defaults to false. Enable it when you need the generated video’s final frame.

Input requirements

Do not directly upload reference images or videos containing real-person faces unless your workflow and account are authorized to use that material.

Images

Property	Requirement
Input methods	Public URL, Base64 data URI, or asset ID
Formats	JPEG, PNG, WebP, BMP, TIFF, GIF, HEIC, HEIF
Aspect ratio	Width/height from 0.4–2.5
Dimensions	Each side from 300–6000 px
File size	Less than 30 MB per image
Count	1 first frame; 2 first/last frames; 1–30 reference images

Videos

Property	Requirement
Input methods	Public URL or asset ID
Containers	MP4 or MOV
Video codecs	H.264/AVC or H.265/HEVC
Audio codecs	AAC or MP3
Resolution	480p or 720p
Duration	2–30 seconds each; no more than 30 seconds total
Aspect ratio	Width/height from 0.4–2.5
Dimensions	Each side from 300–6000 px
Total pixels	409,600–8,295,044 pixels per frame
File size	Up to 200 MB per video
Frame rate	24–60 FPS
Count	Up to 10 reference videos

Audio

Property	Requirement
Input methods	Public URL, Base64 data URI, or asset ID
Formats	WAV or MP3
Duration	2–30 seconds each; no more than 30 seconds total
File size	Up to 15 MB per audio clip
Count	Up to 10 reference audio clips

The complete request body must not exceed 64 MB. Avoid Base64 for large files.

Native multilingual generation

Seedance 2.5 accepts prompts and generates spoken audio in Chinese, English, Spanish, Indonesian, Malay, Thai, Arabic, Portuguese, Vietnamese, Japanese, and Korean. See the Native languages capability example above for its reference image and generated output.

Prompting tips

Organize prompts as subject + action/event + scene/environment + visual style + camera/editing + sound. Omit parts you do not need.
Refer to media as @image1, @video1, and @audio1, and state what each item should and should not contribute.
Distinguish audio intent with punctuation: music in (), sound effects in <>, dialogue in {}, and subtitles in 【】.
For non-Chinese dialogue, state the language before the dialogue.
Use timestamps such as [0–10s], [10–20s], and [20–30s] to structure a long narrative.

Query and retain the result

The create endpoint returns an asynchronous task ID. Poll GET /v1/video/generations/{task_id} until the task reaches SUCCESS, then read result_url.

Task records can be queried for 7 days after submission.
Generated video URLs remain available for 24 hours and allow up to 100 downloads.
Download or transfer the result promptly after success.

Seedance 2.5 API Reference

View complete request and response fields for every supported capability.

Query a Seedance task

Poll task status and retrieve the generated video.

Sources: BytePlus Seedance 2.5 and Volcengine model pricing.

​What’s new

​Capability examples

​Task types and classification

​Select the reference task type explicitly

​Smarter duration and aspect ratio control

​Video editing

​Video extension

​First-frame and first-and-last-frame generation

​Text and reference generation

​Create a task

​Draft to final

​Draft workflow FAQ

​Additional workflows

​Advanced request controls

​Output specifications

​Resolution and pixel dimensions

​Output container

​Audio, watermark, and final frame

​Input requirements

​Images

​Videos

​Audio

​Native multilingual generation

​Prompting tips

​Query and retain the result

Seedance 2.5 API Reference

Query a Seedance task

What’s new

Capability examples

Task types and classification

Select the reference task type explicitly

Smarter duration and aspect ratio control

Video editing

Video extension

First-frame and first-and-last-frame generation

Text and reference generation

Create a task

Draft to final

Draft workflow FAQ

Additional workflows

Advanced request controls

Output specifications

Resolution and pixel dimensions

Output container

Audio, watermark, and final frame

Input requirements

Images

Videos

Audio

Native multilingual generation

Prompting tips

Query and retain the result