Which AI video models accept an input video on Sume?

There are two different things a video can be to a Sume model. A reference video is extra context that guides a new clip, such as motion or style. A source video is the clip being transformed, and the output keeps its length. They are not interchangeable, and a row that supports one often rejects the other with a 400.

The split is advertised in the catalog. GET /v1/videos/models returns supported_input_references per row, and the video-router capability flags say whether the row can run video-to-video.

Which rows take which?

The table summarizes the catalog. Kling Video v3 Pro and Grok Imagine Video 1.5 accept no input video at all; Grok is image-to-video only.

What are the limits on references?

The limits differ per row, so check them before you upload. Wan 3.0 allows up to five reference videos of at most 15 seconds in total, at 16 frames per second or more. Gemini Omni Flash 1.1 allows up to three reference clips, each 3 seconds or shorter. The other rows do not publish a count in the constraints Sume exposes, so test with the catalog before relying on more than one.

On /v1/videos the references go in input_references as video_url entries. On the legacy Video Router wire, source clips use video_url, and Recast and Genjutsu photos use reference_image_urls.

How do I pick?

Want to keep a source clip's timing and change who is in it: Recast or Genjutsu. Want to edit a clip with a prompt: Omni's edit path on the Video Router. Want a new clip that borrows motion from an example: a row with reference videos. For a text-only workflow, Kling is the simpler choice, since it takes no references to get wrong.