Seedance 2.5 reference to video: how to use references

Reference limits

References are public https links, or files you upload with POST /v1/uploads (images up to 30 MB, clips 50 MB, sounds 15 MB). A first frame (image_url) is a different thing: it fixes the opening picture. See image to video.

Name each reference in the prompt

The model reads references in the order you send them. Refer to them by kind and number, and say what each one is for: “the woman in image 1”, “the bottle from image 3”, “copy the camera move of video 1”, “the music follows audio 1”. A reference you do not mention can still bleed into the look, so send only what the shot needs.

# One Idempotency-Key per job, made once: resend with the same $KEY after a timeout, a 429 or a 5xx
KEY=$(uuidgen)
curl https://api.revana.si/v1/generations \
  -H "Authorization: Bearer $REVANA_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: $KEY" \
  -d '{
  "model": "seedance-2.5",
  "input": {
    "prompt": "The chef from image 1 plates the dish from image 2 in the kitchen from image 3. Shot 1: wide of the kitchen. Shot 2: close-up of her hands plating. Shot 3: she slides the plate toward camera and smiles. Pans sizzling, low kitchen chatter.",
    "image_urls": [
      "https://example.com/chef.jpg",
      "https://example.com/dish.jpg",
      "https://example.com/kitchen.jpg"
    ],
    "duration": 15,
    "resolution": "720p",
    "aspect_ratio": "16:9",
    "audio": true
  },
  "max_cost_usd": 5
}'

Reference clips and sounds

A reference clip shows the model a motion or camera move to follow; a sound gives it the rhythm or voice to match. They can be combined with images in one request.

{
  "model": "seedance-2.5",
  "input": {
    "prompt": "Copy the camera move of video 1: a fast dolly through a crowded market, but in a snowy mountain village at dusk. The music follows audio 1.",
    "video_urls": [
      "https://example.com/dolly-shot.mp4"
    ],
    "audio_urls": [
      "https://example.com/track.mp3"
    ],
    "duration": 10,
    "resolution": "720p",
    "audio": true
  },
  "max_cost_usd": 5
}

Edit a video you already have

Send the clip in video_urls and describe only the change. Leave duration and aspect ratio out: the edit keeps the clip’s own.

{
  "model": "seedance-2.5",
  "input": {
    "prompt": "Same shot, but it is night and raining, neon reflections on the street",
    "video_urls": [
      "https://example.com/street.mp4"
    ]
  },
  "max_cost_usd": 5
}

To swap a person, product or place in a clip and keep the same performance and camera, Nemesis is built for exactly that.

What references cost

Reference images do not change the rate: 480p $0.09 · 720p $0.21 · 1080p $0.52 a second. A request with video input is billed at the lower video-input rate ($0.06, $0.15 and $0.34 a second), and the input clips’ seconds are billed with the output: a 10 second 720p clip made from a 10 second reference is $3.00. The free quote endpoint gives the exact price of any request.

Tips for consistent characters and products

• Several angles, one subject. Three or four pictures of the same face (front, three-quarter, profile) hold a character better than one.
• Plain backgrounds for product and character references, so the model does not copy the background too.
• One job per reference. Do not send a picture that shows both the character and the location; split them.
• Keep the cast small. One or two characters stay consistent more reliably than a crowd.
• Draft at 480p ($0.90 for 10 seconds) until the references read right, then render the final.