Seedance
Seedance 2.0 and 2.5 models, media limits, Draft generation, and billing.
Use POST /v1/video-generation/generate. Seedance 2.5 is an additional model; 2.0 Pro and Fast remain available. Omitting model still selects 2.0 Fast.
For authentication, polling, list responses, and errors, see the AI Video overview.
Model limits
| Model | RPM |
|---|---|
doubao-seedance-2-pro | 5 |
doubao-seedance-2-fast | 5 |
doubao-seedance-2-5 | 5 |
| Rule | 2.0 Pro / Fast | 2.5 |
|---|---|---|
| Prompt characters | 3000 | 5000 |
| Output seconds | 4–15 | 4–30 / -1 |
| Reference images | 9 | 30 |
| Reference videos | 3; each 2–15s, total ≤15s | 10; each 2–30s, total ≤30s; edit: each ≥4s |
| Reference audio | 3; image/video also required | 10; each 2–30s, total ≤30s; audio-only allowed |
| Resolution | Pro: 480p / 720p / 1080p; Fast: 480p / 720p | 480p / 720p / 1080p; Draft: 480p |
| Aspect ratio | 16:9, 4:3, 1:1, 3:4, 9:16, 21:9 | 16:9, 4:3, 1:1, 3:4, 9:16, 21:9, adaptive |
Defaults for 2.5: resolution: "720p", ratio: "adaptive", duration: 5, omniReferenceTaskType: "auto", draft: false.
2.5 modes and media metadata
omniReferenceTaskType is auto (default), reference, edit, or extend. Edit and extend require video input and ratio: "adaptive"; edit also requires duration: -1. First-frame and first/last-frame generation require adaptive. adaptive lets the provider choose output geometry; -1 lets it choose duration. These are accepted parameter values, not omitted fields. With multiple videos, identify the target in the prompt.
Use one first_frame and optionally one last_frame. Frame roles cannot mix with reference images, videos, or audio. Reference roles are reference_image, reference_video, and reference_audio.
Put metadata inside each image_url, video_url, or audio_url object alongside url. Each video and audio requires its own duration in seconds. Send actual width, height, size (bytes), and video fps when available; one aggregate duration cannot replace per-file metadata.
Images: each side 300–6000 pixels, aspect ratio 0.4–2.5 (inclusive), size strictly below 30 MiB. Videos: each side 300–6000, aspect ratio 0.4–2.5, area 407,696–8,295,044 pixels, 24–60 fps, up to 200 MiB. Audio: up to 15 MiB. Video and audio totals each have their own 30-second cap.
2.5 reference request
{
"model": "doubao-seedance-2-5",
"content": [
{ "type": "text", "text": "Use Video 1 motion and Audio 1 rhythm" },
{
"type": "video_url",
"role": "reference_video",
"video_url": {
"url": "https://example.com/reference.mp4",
"width": 1280,
"height": 720,
"duration": 5,
"fps": 24,
"size": 8000000
}
},
{
"type": "audio_url",
"role": "reference_audio",
"audio_url": {
"url": "https://example.com/reference.wav",
"duration": 4,
"size": 800000
}
}
],
"omniReferenceTaskType": "reference",
"resolution": "720p",
"ratio": "16:9",
"duration": 10,
"requestId": "reference-request-unique-id"
}Draft → final
Create a playable draft with draft: true and resolution: "480p". Poll its task until success. Within 14 days of creation, request a new 1080p final using the ListenHub taskId as draftTaskId. It must be your own draft in the same application/API identity scope. The API restores its original inputs and channel; omit content and other generation settings from the final request.
{
"model": "doubao-seedance-2-5",
"draftTaskId": "507f1f77bcf86cd799439011",
"requestId": "final-request-unique-id"
}Use a new requestId (8–128 characters) for each intended generation and reuse it for network retries of that same submission. A UUID is suitable. Task details expose optional draft, draftTaskId, and finalTaskIds; draftEpisodeId and finalEpisodeIds may be present for visible generated content.
Estimation and settlement
Draft and final are separate tasks and separate charges. A final does not subtract the draft charge or treat the draft output as another reference video. A failed final refunds its own reserved charge and leaves the successful draft available.
Call POST /v1/video-generation/estimate-credits with the same model, resolution, ratio, duration, omniReferenceTaskType, and draft. Send referenceImages, referenceVideos, and referenceAudios as metadata arrays with role and each file’s metadata, without URLs. To estimate a final, send only model and draftTaskId.
Video estimates use the sum of input-video durations and the selected output geometry at 24 fps, subject to minimum-token rules. Native input dimensions validate each file; they are not independently priced by pixel area. Audio-only requests use the no-video rate; audio seconds are not added to video seconds. Automatic parameters use a conservative estimate, and completion settles against actual provider token usage. The estimate is not a guaranteed maximum.
2.5 has its own rate table. Relative to 2.0 Pro, the per-token rate rises by 52.86% / 48.84% at 480p or 720p (without / with video input), and 51.95% / 48.94% at 1080p. This does not imply the same percentage change in every task or a fixed profit margin.
OpenAPI estimates return tokens and estimatedCostUsd, not internal credits. Reserve the estimate at submission; actual-token settlement can refund or add the difference. Keep the returned task ID and check its state before retrying.
Failures and retries
Missing per-file duration, incompatible frame/reference roles, or an invalid mode/ratio/duration combination fails request validation. For example, edit with duration: 5 is invalid; use adaptive and -1.
An unavailable or inaccessible draft returns 32001; a non-successful, expired, or incompatible draft returns 32004. For example, a draft older than 14 days cannot be finalized. Create a new draft instead.
Upstream moderation or provider failure is reported by the task’s failed state and existing error fields. The 5000-character limit does not guarantee provider acceptance. Do not treat a submission timeout as a new generation: check the existing task or retry with the same request ID.