RelayDance / Guides

How to call the Seedance 2.5 video API (30-second clips, 50 reference assets)

Seedance 2.5 is ByteDance's newest video generation model line, and it is available on RelayDance through the same OpenAI-compatible task API as every other model we carry. No queue, no identity verification. This guide covers what is new in 2.5, the exact model names, per-second pricing, and a copy-paste example.

What is new in 2.5 compared with 2.0

• Clips up to 30 seconds (2.0 tops out at 15). Pass "seconds": "30" explicitly.
• Up to 50 reference assets per request (2.0 allows far fewer): product shots, character sheets, style frames.
• Audio-only references: supply a voice or ambience clip as a reference input.
• .mov input accepted alongside .mp4 for video references.
• 11 prompt languages handled natively, including English, Chinese and Japanese.
• Resolutions: 480p and 720p only. There is no 1080p or 4K tier on 2.5; for those, use Seedance 2.0 Pro 1080p, native 4K, or our 4K upscale endpoint on a 2.5 output.

Models and pricing

Billing is per second of actual output. The API pre-authorizes the worst case when the task is submitted, then settles against the real rendered length and refunds the difference automatically. Failed generations are refunded in full.

Submit a task

curl -X POST https://relaydance.com/v1/video/generations \
  -H "Authorization: Bearer $RELAYDANCE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "doubao-seedance-2-5-720p",
    "prompt": "A chef plating a dessert in a bright kitchen, camera slowly orbiting the dish",
    "seconds": "10",
    "metadata": {
      "generate_audio": true,
      "content": [
        {"type": "image_url", "image_url": {"url": "https://example.com/dish-reference.jpg"}}
      ]
    }
  }'

Poll for the result

curl https://relaydance.com/v1/videos/task_xxxxxxxx \
  -H "Authorization: Bearer $RELAYDANCE_API_KEY" 

Status moves through queued and in_progress to completed, and the response then carries the video URL. Always poll /v1/videos/{id}.

Reference assets and multi-image prompts

Attach references in metadata.content as image_url entries. With several images, refer to them in the prompt as @image1, @image2 and so on, in the order they appear in content. Multi-image reference and first/last-frame control are mutually exclusive. Files you have registered in your RelayDance asset library can be referenced with an asset:// URL placed in the image_url.url field.

Two behaviors worth knowing before production

• Editing, extension, and first/last-frame tasks require adaptive ratio. If the request pins a fixed aspect ratio on those task types, the upstream rejects it. Leave ratio unset (adaptive) for those flows.
• Do not describe background music in the prompt if audio is on. A prompt that asks for a song or BGM can cause the generated soundtrack to fail an output copyright check, which rejects the whole video even though the visuals are fine. Either drop the music wording (ambient sound and speech are unaffected) or pass "generate_audio": false and add music in post. Silent generation also costs less.

When to pick 2.5 vs 2.0

Pick 2.5 for long-form clips (over 15 seconds), heavy reference workflows (dozens of assets), or audio-referenced motion. Stay on 2.0 Pro when you need 1080p or native 4K output, or the lowest cost per second at 720p ($0.190/s on 2.0 Pro vs $0.289/s on 2.5). Both lines run behind the same endpoint, so switching is a one-string change.

Updated 2026-08-29 · API reference · Pricing