Seedance 2.5 with a reference video: Sume prices 15 extra seconds

When a Seedance 2.5 request on Sume includes a reference video, the price estimate adds 15 assumed seconds of input to the output length, then multiplies the result by 0.6. A 5 s 720p clip rises from $2.89 to $6.94. A 30 s 720p clip falls from $17.34 to $15.61. The crossover sits at 22.5 seconds of output.

The rule in the pricing code

Seedance is billed per video token, not per second. Sume's estimator in packages/provider-pricing counts tokens as width x height x seconds x 24 / 1024. When the request carries a reference video (reference_video_urls, video_urls and similar keys), it does two things.

First, it adds SEEDANCE_REFERENCE_VIDEO_INPUT_SECONDS_ASSUMED, which is 15, to the output duration before counting tokens. Sume does not probe the length of your clip for this estimate, so a 3 s reference and a 14 s reference are priced the same. Second, it multiplies the provider cost by 6/10. Then the house margin of 1.25 applies, and the amount rounds up to the cent.

Prices with and without a reference video

These figures use the 16:9 frame sizes in the pricing table (1280x720 at 720p, 1920x1080 at 1080p) and the list rates in the repo: 21,400 micro-dollars per 1,000 tokens, and 23,400 at 1080p. I computed them with the same formula, so treat them as estimates and read usage.cost on the finished job for the charge.

What this means for planning

The extra charge dominates on short clips. For a 5 s motion-reference test, the reference video more than doubles the price at 720p, and it is 2.4x at 1080p. Over about 22 seconds of output the surcharge flips, because 0.6 times the longer token count is below the plain count.

That crossover is a property of the estimate, not a promise about the provider invoice. Do not stretch a clip to 25 seconds to save money. Pick the length the edit needs.

Cheaper ways to get consistency

If you only need a face, product or outfit to stay the same, send reference images instead. The surcharge applies to video references only, and Seedance 2.5 accepts image, video and audio references in input_references. Keep the video reference for motion or camera work you cannot describe in words.

Test the prompt with a short, low-resolution take first. Sume's /v1/videos has no seed (the API returns 400), so a 1080p final is a new take, not an upscale of the draft.

Request shape

A reference video goes in input_references with type video_url. Without frame_images, the API treats the call as reference-to-video.

curl -X POST https://api.sume.com/v1/videos \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "seedance-2.5",
    "prompt": "Match the camera move of the reference, new product on a marble counter",
    "duration": 10,
    "resolution": "720p",
    "aspect_ratio": "16:9",
    "input_references": [
      { "type": "video_url", "video_url": { "url": "https://example.com/move.mp4" } }
    ]
  }'

Check the number before you submit

The reserve is taken when the job is accepted, so an under-funded workspace gets 402 insufficient_credits before anything renders. A 30 s 1080p job with a reference video reserves about $38.39, and the same job without one about $42.65. Compare your balance to the larger of the two if you are unsure which path your client takes.

Different keys feed the estimator: reference_video_urls, video_urls and a few aliases. On /v1/videos the video belongs in input_references with type video_url; the normalizer maps it. If you also send reference images and audio, only the video changes the multiplier, so images and audio add no surcharge in this estimate.