Best AI video model to edit an existing clip: four Sume rows

If you already have a clip and want to change it, only four Sume rows take your clip as the source: gemini-omni-flash-1.1 in edit mode, higgsfield-genjutsu (Motion Transfer), h3-max-recast (person swap) and hailuo-h3-regen (regenerate one H3 clip at 2K). Every other row either makes new video from text, images or references, or takes reference_video_urls as conditioning. Conditioning is not editing: the model uses the clip as a guide, not as the footage it changes. Also remember that sume/auto never routes to Genjutsu or Recast.

What does each editing row do?

The Video Router docs describe video_url as the source clip for these rows only. You cannot combine it with image_url, end_image_url or reference_*_urls.

Which one for which job?

For a prop, color or background change on a short clip, use Omni edit; the API sends no aspect_ratio and the source clip sets the output. For a different person in the same motion, camera and sound, use Recast. For moving a still subject with the motion of a reference video, use Genjutsu. For re-prompting an H3 clip you already made, use H3 Regen, which accepts only a 768p H3 output: at most 1344 by 768, with an audio track, 4.5 to 15 seconds. Genjutsu is listed only where its provider is configured, and the Tencent rows only where the Tencent switch serves them, so check the live catalog first.

What can go wrong?

Omni edit refuses generate_audio: false, bitrate_mode and reference_audio_urls, because the model always writes synced audio. A video_url sent to any other row is a 400. Read the row's constraints from GET /v1/video-router/models/{model_id} before you build a pipeline around one.