Guides · Video Generation

Video Generation

Generate cinematic video clips with Seedance 2.5 — text-to-video, image-to-video, and video-to-video.

Using Portrait as input

Reference virtual or real-person portraits in video generation by passing asset://<asset-id> as multimodal input. asset:// is supported only by seedance-2.0 / seedance-2.0-fast — not seedance-2.5. See Portrait guide for the end-to-end onboarding flow.

Overview

Generate short cinematic video clips from a text prompt, reference images, or reference video using Seedance models. seedance-2.5 is the current flagship and covers text-to-video, image-to-video, and video-to-video (drive a new clip from a reference video via a reference_video).

Creating a video

Submit a video generation task using the Videos API:

cURL

1	curl https://api.alltoken.ai/v1/videos/generations -H "Authorization: Bearer $ALLTOKEN_API_KEY" -H "Content-Type: application/json" -d '{
2	"model": "seedance-2.5",
3	"prompt": "A serene mountain lake at sunrise, cinematic 4K",
4	"ratio": "16:9",
5	"duration": 5,
6	"resolution": "720p"
7	}'

Reference materials

Beyond a text prompt, Seedance can take reference images, video, and audio through the content[] array. Each item has a type and a role:

Material	`type`	Available `role`
Image	`image_url`	`reference_image`, `first_frame`, `last_frame`
Video	`video_url`	`reference_video`
Audio	`audio_url`	`reference_audio`

Two ways to supply each material, and they can be mixed within one request:

Already on a public host — pass the http(s) URL inside the type object, e.g. { "type": "image_url", "image_url": { "url": "https://..." }, "role": "reference_image" }. It must be publicly reachable upstream (no private or signed-only links).
Local file — presign and upload it, then reference the returned upload_id: { "type": "image_url", "upload_id": "upl_...", "role": "reference_image" }.

Warning

The data: base64 inline path was removed on 2026-08-06. A video request carrying data: media now returns 400 inline_media_removed. Migrate by replacing the inline URL with an upload_id — every other field stays the same. http(s) URLs and upload_id are unaffected.

Migrating an inline reference to an upload

1	- { "type": "image_url", "image_url": { "url": "data:image/png;base64,iVBOR..." }, "role": "reference_image" }
2	+ { "type": "image_url", "upload_id": "upl_9f3c...", "role": "reference_image" }

Roles, defaults, and frame rules

You may write type in the short form image / video / audio (the gateway normalizes it), but omitting role applies a default — and for images that default is not what you might expect:

Short `type`	Normalizes to	Default `role` when omitted
`image`	`image_url`	`first_frame` — not `reference_image`!
`video`	`video_url`	`reference_video`
`audio`	`audio_url`	`reference_audio`

So { "type": "image", "upload_id": "upl_..." } is treated as a first frame, not a reference image. Prefer the long form with an explicit role and don't rely on defaults. The GET task response echoes the detected input_type so you can verify how the gateway read your content[].

Warning

Frames and visual references can't be mixed. Upstream treats them as exclusive paradigms: first_frame / last_frame (define the start/end picture) versus reference_image / reference_video (provide style or subject reference). Combining both kinds of visual material in one request returns 400 invalid_request. reference_audio is exempt — "an image as first frame + audio as reference" is a supported combination.

Input constraints

Constraint	Value	Whose limit
Reference video pixel count	≥ 409,600 (i.e. 640×640 and up)	Upstream model — passed through verbatim
Per-file upload size	See the Media Uploads purpose caps	AllToken
Reference image count	Upstream docs say 9; we don't cap it	Upstream

When a reference video is too small, the upstream error is passed through unchanged (for example InvalidParameter: ... video pixel count ... must be greater than or equal to 409600 ...). Note that frames and camera_fixed are not supported by the 2.0 series and are silently ignored rather than rejected.

Supported models

Call GET /videos/models for the live list. seedance-2.5 is the current flagship (text-to-video, image-to-video, and video-to-video); asset:// portrait references stay on the 2.0 line.

Model	Best for	`asset://` portrait reference
`seedance-2.5`	Flagship — text / image / video-to-video	❌ Not supported (portrait is 2.0-only)
`seedance-2.0` / `seedance-2.0-fast`	Portrait (`asset://`) workflows	✅ Supported
`seedance-1.5-pro`	Earlier generation	❌ Not supported (use plain image URLs)

See Portrait guide for how to obtain an asset_id.

Parameters

model — "seedance-2.5" (flagship), "seedance-2.0", "seedance-2.0-fast", or "seedance-1.5-pro"; call GET /videos/models for the live list
prompt — text description of the desired video
ratio — "16:9", "9:16", "4:3", "3:4", "21:9", "1:1", or "adaptive"
duration — length in seconds; -1 lets the model choose automatically
resolution — "480p", "720p", or "1080p"
content — multimodal input array for image-to-video or references, for example { "type": "image_url", "image_url": { "url": "..." }, "role": "first_frame" }. Images containing real human faces must go through Portrait first — pass asset://<asset_id> after onboarding (see Portrait guide); public URLs of real-person images may be rejected upstream for deepfake / consent compliance.

Checking task status

Video generation is asynchronous. Poll for status:

cURL

$curl https://api.alltoken.ai/v1/videos/generations/{task_id} -H "Authorization: Bearer $ALLTOKEN_API_KEY"

The response includes status (queued, processing, completed, failed, expired, or cancelled) and a download URL when complete.

Cancel a task

Cancel a queued or processing task:

cURL

$curl -X POST https://api.alltoken.ai/v1/videos/generations/{task_id}/cancel   -H "Authorization: Bearer $ALLTOKEN_API_KEY"

Tasks already completed, failed, or cancelled cannot be cancelled.