Seedance 2.5 references: ByteDance says 30 images, Sume allows 9

ByteDance's Seedance 2.5 announcement talks about referencing up to 30 images, 10 video clips and 10 audio clips. On Sume the API validates a Seedance 2.5 job at 9 images, 3 videos and 3 audios, and no more than 12 reference files in total.

Treat the vendor figure as what the model family can take and the Sume figure as what one request to this API accepts.

Side by side

What to do with a big cast

A request over the cap fails validation instead of being silently trimmed, so you see the problem before you are charged.

• Pick the 9 images that carry identity: faces, product, location.
• Use video references for motion or camera, not for every shot.
• Split a long story into jobs and keep the same cast images in each.

Where to confirm

The Video Router docs say to read capabilities from the catalog because each model has a different envelope. For the all-model view, see the reference limits per model.

Why the numbers differ

A vendor limit describes what the model or its first-party app can take. An API gateway adds its own per-request checks so a job is predictable to price and to run. Sume enforces 9 images, 3 videos and 3 audios and 12 files in total for Seedance 2.5 before any money is reserved.

ByteDance's page also says API access was coming soon through BytePlus ModelArk at launch, and that the model was live in Jimeng AI and Doubao Pro. Sume offers it through its own API today.

Workarounds that keep the story intact

Reference limits shape how you plan a series. Rather than push 30 faces into one job, assign each scene its cast of up to 9 images and reuse the hero images across jobs.

For motion, trim the reference clip to the useful seconds first; Sume limits video references to 3 per Seedance 2.5 job.

• Name your reference images by role (hero, product, location).
• Re-use the same URLs for continuity.
• Count files before you submit: 9 + 3 + 3 is more than 12.