# Video to Video

Video to video works the way image to image does. In the chat bar's **Video** category, selecting a clip on the board switches the bar to video to video; dropping the clip switches it back to image to video. There is no mode to pick: what is selected decides.

In the **Image** category a selected clip is still sent as a still frame of it, to the image models.

## What the selection makes

- **One clip alone** — the clip is the input. Auto picks **Kling O3 Standard Edit**: the clip comes back changed to your prompt (restyled, something swapped, a different time of day) with its motion and length kept.
- **A clip with anything beside it** (an image, or more clips) — Auto picks **Seedance 2 Reference**, which makes a new clip from all of them. Up to 3 clips.
- **One image and one clip** — pick **Kling v3 Standard Motion Control** or **Kling v3 Pro Motion Control** in the model list to make the character in the image move the way the person in the clip moves.

The selected clip shows above the prompt as a tile marked V1. The models that cannot take what is selected wait in collapsed buckets below the list ("Needs image input", "Needs video input", "Takes fewer inputs"); open one to see them greyed, with the reason: "One clip only", "Add an image or another clip", "Does not take a video". With nothing or only images selected in Video, the video-to-video models are under "Needs video input".

## The models

Edits (one clip, returned at its own length):
- **Kling O3 Edit** (Standard, the default, and Pro) — clips of 3 to 15 seconds, at least 720 pixels on each side, MP4 or MOV. Keeps the clip's sound unless switched off.
- **Kling O1 Edit** (Standard and Pro) — clips of 3 to 10 seconds, 720 to 2160 pixels on each side.
- **Gemini Omni Flash 1.1 Edit** — clips up to 10 seconds; output at 360p, 720p, 1080p or 4K. Simple instructions work best: "Make it night. Keep everything else the same."
- **Wan 2.7 Edit** — clips of 2 to 10 seconds, at most 100 MB; 720p or 1080p; keeps the clip's sound or lets the model make new sound.
- **LTX 2.3 Retake** — redoes one section of a clip up to 20 seconds long and keeps the rest. Set the section with the Start and Length pills; it must fit inside the clip.
- **LTX 2.3 Extend** — adds 2 to 19 seconds after or before the clip, continuing from the seconds of it it reads (Context). Added seconds and context together are at most 20.

New clips guided by a clip:
- **Kling O3 Reference** (Standard and Pro) and **Kling O1 Reference** (Standard and Pro) — a new clip that borrows the clip's camera, motion and style, with up to 4 reference images, at a length you choose (O3: 3 to 15 seconds, O1: 3 to 10). Call the clip @Video1 and the images @Image1, @Image2 in the prompt.
- **Seedance 2 Reference** and **Seedance 2.5 Reference** — images and clips together as references for one new clip.
- **Kling v3 Standard Motion Control** and **Kling v3 Pro Motion Control** — one character image and one motion clip. Facing "As in the clip" takes a clip up to 30 seconds; "As in the image" up to 10. The prompt is optional.

## What it costs

- An edit, a retake and Motion Control are charged per second of the clip you send. The model list shows them as a rate, such as "27/s", and the price beside Send appears once the clip's length is known, which for a clip just made is a moment after it lands.
- An extension is charged for the seconds it adds plus the context seconds it reads.
- A Kling Reference clip is charged for the length you set, like any video.
- A clip longer than the model takes is refused before any credit is spent.

## Where the result goes

The result lands on the board beside the clip it came from, joined to it by a line, and in your video library. Its preview is made from its own frames once it arrives, so it can read "No preview" for a moment.

## In workflows and with the assistant

- A [workflow](/docs/product/workflows) can contain a video-to-video step and can start from a clip: run it with a clip selected and the clip is the input.
- The assistant has no video-to-video tool. It can explain these models and show you how, but the chat bar runs them.