Doubao Seedance 2.5

Models / Tiers

The second row is Seedance 2.0 Official, for comparison: 2.5 at 720p costs about 37% more, and buys 30-second clips, 50 reference assets and the edit / extend task types. Upstream bills on generation tokens, so the table shows derived per-second values; the charge follows the tokens actually produced. Charged only on success, refunded on failure.

Aliases: seedance-2-5 / seedance-2.5-official / doubao-seedance-2-5-260628 all point at the same model; the canonical name echoed in the response is seedance-2.5.

Authentication

All requests carry the API key in the header:

Authorization: Bearer YOUR_API_KEY

Video Generation

# Text-to-video (model: seedance-2.5)
curl -X POST https://api.apimodels.app/v1/video/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "seedance-2.5",
    "prompt": "a hummingbird hovering near a bright flower, slow motion",
    "resolution": "720p",
    "aspect_ratio": "16:9",
    "duration": 8
  }'

# Poll until completed (result files are kept 7 days)
curl "https://api.apimodels.app/v1/video/generations?task_id=TASK_ID" \
  -H "Authorization: Bearer YOUR_API_KEY"

# "duration": -1  → the model picks the best length for the prompt.
# "output_format": "mov"  → H.264 + yuv444p + PCM instead of the default mp4.
#
# Real person: raw real-person images/videos are accepted — on an upstream
# moderation rejection we auto-register the material and retry transparently.
# To REUSE a fixed person across tasks, register it once as asset://:
#   "reference_image_urls": ["asset://asset-2026...-xxxx"]

Parameters

Reference asset limits

One job takes up to 50 reference assets: 30 images + 10 videos + 10 audios. Assets are cited positionally in the prompt — the first entry of reference_image_urls is 图片1 / @image1, the first of reference_video_urls is 视频1 / @video1, and audios follow the same pattern. Telling each asset what to supply and what NOT to supply beats piling more of them on.

On 2.5 you may pass audio as the only reference (a voice track with no image, say); 2.0 requires audio to come with an image or a video.

Reference material hard specs

Beyond counts and durations, upstream hard-validates the media files themselves — out-of-spec input fails the task outright. The most common trap is frame rate: 120fps clips (phone slow-motion, screen recordings) are rejected; re-encode to 30fps first.

Request bodies with Base64 payloads are capped at 64 MB — pass large files as public URLs instead of inlining them. Out-of-range frame rates are validated locally at create-time with a clear 400, rather than failing mid-task.

Real-person assets (asset://)

Real-person reference images/videos can be passed as raw URLs — if upstream moderation rejects one, we auto-register it in the asset library and retry transparently (a few seconds of extra latency). To reuse the same person across tasks, or to skip that automatic step, register an asset:// reference once: it lasts, costs $0.01 per successful registration, and failures (HTTP 422) are never charged.

# 1) Create an asset group (once). Seedance 2.5 uses the DEFAULT asset library --
#    do NOT pass "account": "intl" here (that is a separate library for the
#    Dreamina international line; the two do not interoperate).
curl -X POST https://api.apimodels.app/v1/assets/groups \
  -H "Authorization: Bearer YOUR_API_KEY" -H "Content-Type: application/json" \
  -d '{"name": "my-people"}'
# -> { "data": { "id": 1, ... } }

# 2) Register the face from a public URL. The call waits until the asset is
#    confirmed usable (~3s). Charged $0.01 ONLY on success -- an unfetchable or
#    moderation-rejected image returns HTTP 422 and is NOT charged.
curl -X POST https://api.apimodels.app/v1/assets \
  -H "Authorization: Bearer YOUR_API_KEY" -H "Content-Type: application/json" \
  -d '{"url": "https://example.com/face.jpg", "asset_type": "Image", "group_id": 1}'
# -> { "data": { "id": "asset-2026...-xxxx", "asset_url": "asset://asset-2026...-xxxx", "status": "Active" } }

# 3) Use asset_url in the seedance-2.5 request -- reference_image_urls to borrow
#    the person's look, or first_frame_url to animate exactly that person.

• Seedance 2.5 uses the default asset library: do not pass an account field when creating the group. account: "intl" is a separate library for the Dreamina international line and the two do not interoperate — for that flow see Dreamina Seedance 2.0 (Intl).
• To animate a specific person, put the asset:// in first_frame_url; to borrow looks/outfit only, put it in reference_image_urls.
• Asset management: POST /v1/assets/list (paginated list), POST /v1/assets/get (fetch one), POST /v1/assets/delete, POST /v1/assets/update (rename).

Task types and their hard constraints

2.5 adds video editing (replace / add / remove objects in frame, and edit the audio track) and video extension (continue the clip forwards or backwards, round after round). Those two, plus first-frame and first+last-frame, all lock the output ratio — the result follows the source and you cannot override it.

• "Forced" means anything else is an error. A video edit always follows the source clip ratio and length (the result may come back up to 0.4s shorter).
• Upstream reports these conflicts asynchronously — the task queues first and only later comes back with InvalidParameter.TaskTypeConstraint. We reject them at create time instead and return 400 immediately, so you find out on the spot.
• The source video for an edit must itself be 4-30 seconds.

Video editing and video extension

Both task types share one endpoint: put the source clip in reference_video_urls, refer to it in the prompt as 视频1 / @video1, and spell out what to change or what to continue. task_type is optional (the backend infers it from the prompt) but stating it outright is more reliable.

# ── Video EDITING — replace / add / remove objects, or edit the audio track ──
# The source clip goes in reference_video_urls and is cited as 视频1 / @video1.
# Editing forces aspect_ratio "adaptive" AND duration -1 (the output keeps the
# source ratio and length). Send either one and you get an immediate 400.
curl -X POST https://api.apimodels.app/v1/video/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "seedance-2.5",
    "prompt": "把视频1改成夜晚雨景,保留人物动作和镜头运动",
    "reference_video_urls": ["https://example.com/source.mp4"],
    "task_type": "edit"
  }'

# ── Video EXTENSION — continue the clip forwards or backwards, round after round ──
# Extension also forces aspect_ratio "adaptive", but duration is yours: 4-30 or -1.
curl -X POST https://api.apimodels.app/v1/video/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "seedance-2.5",
    "prompt": "续写视频1,人物继续向前走进地铁站",
    "reference_video_urls": ["https://example.com/source.mp4"],
    "duration": 10,
    "task_type": "extend"
  }'

Billing

Billed on real upstream token usage (usage.completion_tokens), charged only on success and refunded on failure. Volcengine computes it as:

tokens = (input video seconds + output seconds) × width × height × fps ÷ 1024

The per-second prices in the tier table (480p $0.120, 720p $0.270) are display values derived from that formula; the real bill follows the tokens actually produced.

Notes

• Async: create returns a taskId; poll GET /v1/video/generations?task_id= until completed. Result files are kept 7 days — download or re-host promptly.
• first_frame_url / last_frame_url and reference media (reference_image_urls / reference_video_urls / reference_audio_urls) are mutually exclusive — one input mode per request. Sending both returns a 400 rather than silently ignoring one side. To animate a specific real person, put that portrait’s asset:// straight into first_frame_url.
• Real-person material (images or videos) can be passed raw — an upstream rejection triggers an automatic asset-library retry. To reuse a fixed person, register an asset:// (see “Real-person assets” above).
• Native synced audio (voice / SFX / music), the web search tool and the asset:// real-person library all work exactly as they do on 2.0.
• Prompts are natively multilingual: Chinese, English, Spanish, Indonesian, Malay, Thai, Arabic, Portuguese, Vietnamese, Japanese and Korean. There is no need to translate your prompt into English first.
• mov is for post-production (H.264 + yuv444p + PCM, higher colour fidelity for grading and keying); if the clip goes straight to a web page or a social platform, keep the default mp4 — some players cannot open mov.
• Use real-person material only with subjects you have the rights/consent to depict; portraits and outputs are screened by upstream moderation.

Try it in the Playground