video_url or reference_video_urls? Which Sume models edit a clip
Sume has two ways to pass a video: video_url is a source to edit or recast, reference_video_urls only conditions a new clip. Which models take which.

Two fields that look alike
On Sume's Video Router, a clip can arrive as video_url or as reference_video_urls. They do different jobs. video_url is the source: the output is built from that clip and keeps its length. reference_video_urls is evidence for a fresh generation, a look or motion to follow, and the model still writes its own clip.
The Video Router docs say it plainly for Gemini Omni Flash 1.1: video_url is the edit source, not a reference, and it cannot be combined with image_url, end_image_url or the reference_*_urls fields.
Who takes a source clip
Only three catalog rows take video_url, according to Sume's Video Router notes in the repository. Each one is an explicit pick; the Auto router does not route to Genjutsu or Recast.
| Model id | Source clip | Other input | Length |
|---|---|---|---|
| gemini-omni-flash-1.1 | video_url, edit mode | Prompt describes the edit | Output follows the source |
| higgsfield-genjutsu | video_url, Motion Transfer | 1 to 8 reference images | 4 to 30 seconds |
| h3-max-recast | video_url, person swap | 1 to 4 reference photos, prompt optional | 5 to 30 seconds |
Who takes a reference clip
The Videos API docs say video and audio references are honored by the Seedance 2.x models, Wan 3.0, MiniMax H3 and MiniMax H3 Max. Gemini Omni Flash 1.1, Genjutsu and Recast accept video references but not audio. Gemini's reference clips are capped at three, each three seconds or shorter.
So a short motion reference on Seedance does not turn into a recast of that clip. If you want the same footage with different people or props, you want a source row, not a reference row.
A quick way to choose
- Keep the footage, change a product or object: Gemini Omni Flash 1.1 edit, with a prompt that names the change.
- Keep the motion, swap the characters: Genjutsu, with up to 8 images.
- Keep motion, camera, cuts and sound, swap the people: Recast, with up to 4 photos.
- Use a clip as inspiration for something new: a reference video on a Seedance or MiniMax model.
Sources
Related posts
More in Developers
- VideoObject for a product video: thumbnail, date and duration
Google's VideoObject needs name, thumbnailUrl and uploadDate. How to fill them and the duration from a Sume trim result and a video-frames still.
- A /v1/videos id is a Sume job id: poll /v1/jobs for more (Python)
POST /v1/videos returns an id that is also a Sume job id. Poll GET /v1/jobs/{id}/status for terminal and next_poll_after_seconds, and cancel with /v1/jobs.
- Virtual try-on API: which Sume call returns an image, which a video
Need a try-on photo or a try-on clip? On Sume the two catalog try-on Formats return video; a still comes from the image API. The table, plus one call for each.
- Unit-test a Sume submit-and-poll loop in Vitest with fake replies
Test a Sume polling loop without spending credits: inject the sleep, replay queued then completed replies, and assert next_poll_after_seconds is honored.
Written by Sume