A 17.3 s dance source on Genjutsu bills 18 s: $7.16 at 480p

On Sume, Genjutsu Motion Transfer (higgsfield-genjutsu) wants the duration to equal the inspected length of the source video, rounded up. A 17.3-second source is submitted as 18 seconds and bills $7.16 at 480p or $15.33 at 720p. The supported range is 4 to 30 seconds.

The rule from the catalog

The catalog says: one source video and 1 to 8 reference images are required; there is no text-only generation; the default is 480p; and duration must be the inspected input video's length in seconds, rounded up (4 to 30). The output preserves the source length and framing, so there is no aspect_ratio. The model is in the catalog only when its provider is configured.

• Request on /v1/videos: input_references with exactly one video_url and 1 to 8 image_url entries.
• Request on /v1/video-router/generate: video_url plus reference_image_urls.
• Not accepted: aspect_ratio, frame_images, generate_audio, bitrate_mode, audio references.

Price at 17.3 seconds

Provider list per second is $0.318 at 480p and $0.681 at 720p. Sume bills list times 1.25 and rounds up to the cent. The billable length is the rounded-up 18 seconds.

Measure first

Do not guess the source length. Run video inspect with frames: false and read the probe duration, which costs only the probe compute. Round it up yourself: ceil(17.3) = 18. The same step catches a source over 30 seconds before you pay for a refusal.

A source shorter than 4 seconds is below the supported range, so probe before you submit.

Fit for a vertical post

Because the output keeps the source framing, a 9:16 dance source gives a 9:16 result. A 16:9 source gives 16:9, and Sume will not reframe it. If your target is a TikTok ad, the specification page lists 9:16 at 540x960 minimum, 16:9 at 960x540 and 1:1 at 640x640. Probe the 480p output before you rely on that floor.

Motion Transfer does not make sound (audio is false in the catalog). Add the sound at a Timeline step, where the render is $0.10 per started minute.

Why the length rounds up

Genjutsu Motion Transfer bills for the inspected input video's length rounded up to a whole second, within 4 to 30 seconds, so a 17.3 s dance is billed as 18 seconds. Probe the clip first with video inspect to read the real duration, and send that rounded value as the job duration. Sending a shorter duration than the source is the common mistake: the docs say the duration must be the inspected length rounded up.

If the source is under 4 seconds, the minimum applies, so a 2-second move is billed as 4. A source over 30 seconds is out of range and must be trimmed first, which costs $0.02 per trim job.

Arithmetic for the two resolutions

At 480p the list is 318,000 micro-dollars per second, or $0.318. Times 18 seconds is $5.724 list, times 1.25 is $7.155, rounded up to the cent is $7.16. At 720p the list is $0.681 per second: 18 x 0.681 = $12.258 list, times 1.25 = $15.3225, rounded up to $15.33.

The 720p job costs 15.33 / 7.16 = 2.14 times the 480p one, so check the 480p result before you pay for 720p. TikTok's non-Spark minimum for 9:16 is 540x960, so confirm the pixel size of the output with a probe before using it as an ad.