Keep one character consistent in a 30-second Seedance 2.5 clip

For a character that must look the same for 30 seconds, send two to four clean reference images in reference_image_urls on a seedance-2.5 request instead of describing the character in words only. One pass of up to 30 seconds also removes the seams where identity can drift between stitched short clips. Sume does not guarantee consistency, so test it on a cheap 480p draft first.

What the vendors say, and what Sume accepts

ByteDance's Seedance 2.0 page says the model supports images, audio and video as references and gives control over performance, lighting, shadow and camera movement (ByteDance Seed: Seedance 2.0). BytePlus says Seedance 2.5 accepts up to 50 reference assets: 30 images, 10 videos and 10 audio files (BytePlus: What is Seedance 2.5).

Sume's Video Router is stricter than that vendor ceiling. In the schema read on 2026-10-04, a Seedance request takes at most 9 reference images, 3 reference videos, 3 reference audio files, and 12 references in total. The vendor ceiling is not the Sume ceiling.

Choose the references that carry identity

Identity comes from the images, so pick them deliberately. Use the same character at a neutral angle, a three-quarter angle and one full-body view, shot in consistent lighting. Avoid collages, heavy filters and images where the character is small in the frame.

Real photographs of people raise separate questions about rights and provider rules. Read Seedance and real human faces before you send a photo of a real person.

A cheap consistency test

Run the same prompt and references three times at 480p and 4 seconds, which is the shortest duration seedance-2.5 accepts. Each probe costs about $1.07 on Sume's rate card, so three takes cost about $3.22. Compare one still per take; if the face or outfit changes across takes at 4 seconds, a 30-second clip will not repair it.

Then run the real 30-second request. Reference images and first or last frame fields cannot be mixed: the Video Router docs describe image-to-video and reference-to-video as separate shapes.

curl -X POST https://api.sume.com/v1/video-router/generate \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: character-test-001" \
  -d '{
    "model": "seedance-2.5",
    "prompt": "The courier from the reference images walks through a market, then stops to read a parcel label.",
    "reference_image_urls": [
      "https://example.com/courier-front.png",
      "https://example.com/courier-side.png"
    ],
    "resolution": "480p",
    "duration": 4,
    "mode": "async"
  }'

Judge the result honestly

Extract stills at the start, middle and end of the clip with video frames (the clip must already be stored on media.sume.com, so import a downloaded MP4 first) and put them side by side with your references. Judge outfit, hair, and proportions, not only the face. If the test passes, repeat at 720p; if it fails, change the references before changing the prompt.