Seedance 2.5 takes 30 image references per pass: a shot-list ad

ByteDance says Seedance 2.5 takes up to 30 images, 10 video clips and 10 audio clips in one pass, and generates up to 30 seconds (read 2026-10-05). On Sume you call it as seedance-2.5 in the Video Router, which accepts 4 to 30 seconds at 480p, 720p or 1080p. The way to use that many references in an ad is to turn the shot list into labelled references and a prompt that names each one, then check how many references Sume lets you attach before you go past a handful.

From shot list to reference set

A shot list has rows like: hero pack, close-up of the texture, a hand opening the lid, the kitchen, the end card. Give each row at most one or two images. A set of ten pictures that each do one job beats thirty that repeat. Put the same order in the prompt, so the text and the images line up.

• Product views: front, side, top, label close-up.
• Setting: one or two environment photos, no product in them.
• Hands or people: only if the shot needs them, with a clear face.
• Audio: a music bed or a voice line, if you want it in the pass.
• Style: one frame from your last campaign, so the grade matches.

The call

Seedance 2.x models accept audio and video references as well as images, per the Video generation page. Send each category in its own field, and keep the prompt about camera and action. Add an Idempotency-Key so a retry cannot double bill.

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: seedance-shotlist-001" \
  -d '{
    "model": "seedance-2.5",
    "prompt": "Shot 1 pack on a shelf, shot 2 texture close-up, shot 3 hand opens lid, shot 4 end card",
    "reference_image_urls": [
      "https://example.com/pack-front.jpg",
      "https://example.com/texture.jpg",
      "https://example.com/hand-lid.jpg"
    ],
    "resolution": "720p",
    "duration": 15,
    "aspect_ratio": "9:16",
    "mode": "async"
  }'

Limits: vendor versus Sume

The vendor page also says the API is coming soon through BytePlus ModelArk and that the model is rolling out in Jimeng AI and Doubao Pro. That is the vendor's own channel. Sume lists seedance-2.5 in its catalog today, so the Sume limits are the ones to read from GET /v1/video-router/models/seedance-2.5.

Start with three to six references and a 12 to 15 second clip. Add references only when a shot comes out wrong for a lack of information, and keep the one that fixed it.

What can go wrong

Two failure modes are common with large reference sets. First, a reference that conflicts with the prompt. If the photo shows a blue bottle and the prompt says green, the result is unpredictable, so write prompts that agree with the images. Second, an end card. A model that sees a logo in a reference may place it in the middle of the clip. If you need the logo only at the end, leave it out of the references and add it afterwards in a Timeline compose or a separate logo clip.

Keep a log of which references went into which clip, next to the Idempotency-Key you used. When a client asks why the lid looks different in clip four, you can answer from the log and not from memory. Every retry is a new paid job, so the tuning goal is fewer retries.