Seedance 2.5 omni reference: how @image, @video and @audio tags work

Omni reference in Seedance 2.5 means you attach images, videos and audio, then name each one in the prompt with a tag such as @Image 1, @Video 1 or @Audio 1. On Sume, write @image1, @video1 or @audio1 in the prompt and the API normalizes them to @Image1, @Video1, @Audio1 before the job runs.

The tag syntax and the special characters come from the ModelArk tutorial; the recommendations come from the prompt guide.

What do the tags refer to?

Each tag names one attached reference. Say what the model should take from it: the face, the outfit, the camera move or the rhythm. A tag with no instruction leaves the model to guess.

What special characters does the tutorial list?

The tutorial defines four bracket styles for sound and text in the prompt.

What prompt structure does the vendor recommend?

The tutorial's formula is subject, action, scene, style, camera and sound. Its prompt guide suggests 1 to 5 subjects and 5 to 10 seconds for reference audio or video.

{
  "model": "seedance-2.5",
  "prompt": "@image1 walks through the market from @image2, handheld camera, warm light, crowd murmur",
  "duration": 8,
  "resolution": "720p",
  "aspect_ratio": "9:16",
  "input_references": [
    {"type": "image_url", "image_url": {"url": "https://example.com/hero.png"}},
    {"type": "image_url", "image_url": {"url": "https://example.com/market.png"}}
  ]
}