Seedance reference video: 15 assumed seconds, 0.6x, real cost

Sending a reference video to Seedance on Sume prices as 8 s plus 15 assumed input seconds at 0.6x. For 8 s at 720p that is $5.22 on seedance-2 against $3.02.

5 min readSume
All posts

Adding one reference_video_urls entry to a Seedance request changes the price estimate even if you also send no extra images. Sume's pricing code counts the output seconds plus 15 assumed seconds of reference input, then multiplies the token cost by 0.6. For an 8-second 720p clip on seedance-2 that moves the Sume price from about $3.02 to about $5.22, a 1.73x jump. The 15 seconds is a constant, not a measurement of your file.

How the estimate is built

Seedance bills video tokens: width x height x seconds x 24 / 1024. With a reference video present, the billable seconds become the requested duration plus 15. The resulting list cost is then multiplied by 6/10. Sume prices the provider list times 1.25 on top, as the docs state for video models.

For 720p 16:9 and 8 s, that is 496,800 tokens instead of 172,800 before the 0.6 factor. The same arithmetic holds for each tier; only the per-1,000-token rate changes.

8 s 720p Seedance, without and with one reference video, Sume price (read 2026-10-07)
ModelNo reference videoWith reference videoRatio
seedance-2-mini$1.51$2.611.72x
seedance-2-fast$2.42$4.171.73x
seedance-2$3.02$5.221.72x
seedance-2.5$4.62$7.971.72x

What this means for a Sora migration

Teams that kept a library of Sora clips as style anchors will want to pass them as references. The video-reference field exists on the Seedance 2.x models, Wan 3.0 and the MiniMax H3 family, per the docs, and Gemini Omni Flash 1.1 takes video references of at most 3 seconds each. On Seedance the price effect is the one in the table.

If the look you want can be carried by one or two still frames, frame_images or image references cost nothing extra in this formula. That is a cheaper first experiment than a reference video.

Short references still pay the full assumption

Because the constant is 15 seconds, a 3-second reference clip is estimated like a 15-second one. Trimming the reference to its useful seconds improves the output; it does not lower this estimate. If cost matters, test whether an image reference gives the same result before you pay for a video one.

Treat the figure as the reserve Sume sets at submit. The poll response's usage.cost is the billable amount, and the docs say reserve and charge both follow list x 1.25.

Checks before a batch

  • Submit one job with the reference video and one without, same prompt and duration, and compare usage.cost.
  • Confirm the model lists video_url in supported_input_references at GET /v1/videos/models.
  • Remember you may send 1 to 3 reference videos on the legacy Video 1.0 route; per-model caps are in the catalog.

A budget line for 40 referenced clips

At the seedance-2 price in the table, 40 clips of 8 seconds at 720p with a reference video come to about $209 (40 x $5.22), against about $121 without references (40 x $3.02). The difference of roughly $88 is the cost of the reference video feature across the batch. Whether it is worth paying depends on how many first-try results you save, which only a small A/B run can tell you.

A tidy way to run that test is to submit the same prompt twice per sample, with and without the reference, and read both usage.cost values from the poll responses. Keep the prompt, duration, resolution and aspect ratio identical so that cost and quality differences come from the reference alone.

Sources

Related posts

More in Pricing

All Pricing posts

Written by Sume