Which AI Video Models Can Edit an Existing Clip? V2V Options

Video-to-video means different things on different pages. Some models restyle a clip by instruction, some swap the person but keep the motion, and some only use a clip as a reference for a fresh generation. Here is what each vendor and Sume document on 2026-10-03.

Instruction-based edit

Google's Omni documentation describes conversational editing through the Interactions API, with the caveat that audio in reference videos is ignored and voice editing is unsupported. On Sume, gemini-omni-flash-1.1 exposes its edit mode through the Video Router video_url field. The edit path cannot be combined with image or reference fields.

Elsewhere: Luma lists Modify Video for Ray 3.2 with clips up to 20 seconds depending on frame rate; Runway's pricing page lists Aleph 2 at 28 credits with a 56-credit minimum; and xAI's guide lists video-edit input for grok-imagine-video-1.5. Sume's Grok row, however, is image-to-video only.

Person swap and motion transfer

Two Sume routes take a source clip and keep its movement. h3-max-recast takes a source video of 5 to 30 seconds, with no shot over 15 seconds, plus 1 to 4 person photos; it keeps motion, camera, cuts and soundtrack, and the prompt is optional. higgsfield-genjutsu takes one source video plus 1 to 8 reference images for 4 to 30 seconds and is listed only when its provider is configured.

Choosing

If you want to change wording, lighting or a prop with a sentence, use the instruction edit. If the choreography is right and the person is wrong, Recast is built for that and is priced accordingly at $0.375 per second. If you want a reference-driven remake rather than an edit, a fresh generation with video references is usually cheaper.

Sume's docs do not publish a quality ranking between these routes, so run a 5-second test on your own source before buying the full 30. See which model fits the inputs you have.