video_url on Sume video: only Omni edit, Genjutsu and Recast take it
On the Sume video API, video_url is a source clip, taken by gemini-omni-flash-1.1 (edit), higgsfield-genjutsu and h3-max-recast. Others use reference_video_u...

On Sume's Video Router, video_url means "this is the clip I want edited or transformed", and only three ids accept it: gemini-omni-flash-1.1 in edit mode, higgsfield-genjutsu for Motion Transfer, and h3-max-recast for person swaps. Every other id that can use a video takes it as reference_video_urls, which conditions a new generation instead.
Mixing the two is the most common mistake. A source clip and a reference clip are different jobs with different pricing.
Source versus reference
The distinction decides both the field name and the billing basis.
| You want | Field | Ids |
|---|---|---|
| Edit this clip with a prompt | video_url | gemini-omni-flash-1.1 |
| Move this motion onto new subjects | video_url plus 1 to 8 reference images | higgsfield-genjutsu |
| Swap the people in this clip | video_url plus 1 to 4 reference images | h3-max-recast |
| Use a clip as a style or content reference | reference_video_urls | seedance 2.x, wan-3.0, minimax-h3 ids, omni |
Rules that follow
You cannot combine video_url with image_url, end_image_url or reference_*_urls in the same request. Edit mode sends no aspect_ratio and the source clip sets the output. If you give a duration with an Omni edit, it is only a reserve hint, with a default of 8 seconds, because the provider does not return an output length. Genjutsu and Recast bill the source length rounded up.
sume/auto never routes to Genjutsu or Recast. Callers select them explicitly.
Picking quickly
Ask what should stay the same. If the whole clip should stay and one thing change, use Omni edit. If the motion should stay and the person change, use Genjutsu or Recast. If nothing in the clip needs to survive and you only want its feel, use a reference video on a generation id. And if the id you picked is not in that list, video_url will fail, so read the catalog row first.
Common errors
Three mistakes account for most video_url failures.
Genjutsu appears in the catalog only when its provider is configured, so a list call on one environment may not show it. Read the list you are actually calling.
- Sending
video_urlto a model that does not list it. - Sending it together with
image_urlorreference_*_urls. - Sending a source longer than the model allows: 4 to 30 seconds for Genjutsu, 5 to 30 for Recast, with no Recast shot past 15.
Sources
Related posts
More in Developers
- Voice one script in six languages: Sume TTS loop and 409 guard
MAI-Voice-2.1 sells one voice across 23 languages. On Sume, set language per line, handle the 409 voice-language guard and price a six-language batch.
- Voiceover for shorts: one sentence per shot using TTS sentence slices
Ask TTS for timestamps.words and sentence segmentation and Sume returns gapless per-sentence slices, so each shot in a short gets its own audio file.
- VoltAgent MCP 30000 ms default timeout vs Sume jobs_wait
VoltAgent's MCP timeout defaults to 30000 ms; Sume's jobs_wait holds 50 to 55 seconds. Raise it past 55000 and re-wait on wait_slice_expired.
- waitForJob on a failed Sume image job: read the record, skip /result
waitForJob resolves for failed and canceled jobs instead of throwing, and /result returns 409 job_not_completed for them. How to branch on job.status.
Written by Sume