Kling 4.0 Video Edit: Change the Shot, Keep the Take

The take is good. The timing works, the camera move lands, the actor hits the mark on the third beat. Then the client asks for the jacket in navy. Or the brand wants the bottle on the counter to be the new bottle. Or the whole thing should feel like it was shot at dusk in Lisbon instead of noon in a car park. A year ago that meant a reshoot.

Kling 4.0 Video Edit is built for exactly that moment. You hand it the clip you already have, tell it what to change, and optionally show it what the change should look like. The motion stays yours. The contents of the frame are up for negotiation. It went live on each::labs on October 24, and this guide is the practical part of our Kling 4 series: every input, how to wire references into your prompt, a workflow you can repeat, and edit prompts you can copy.

What Kling 4.0 Video Edit Actually Does

Kling released Kling 4.0 on October 15, and the headline in Kling's launch announcement is control. Omni Reference accepts up to 15 multimodal references (images, video and audio among them) plus up to 10 keyframes, with the stated goal of keeping characters consistent across shots. Kling also says reference videos can guide performance, action, camera movement and pacing, and that big motion and complex camera work (push-ins, pans, tracking shots, orbits) hold together better than before.

Video Edit is that reference system pointed at footage that already exists. It's a source-video workflow, not text-only generation: the request requires a prompt and a source clip, and everything else is optional steering. That's a different job from generating. When you generate, you're describing a world and hoping the motion comes out right. When you edit, the motion is already right, and you're only describing the difference.

That difference is the whole skill. Most bad AI video editing comes from prompts that describe the entire scene again, which invites the model to reinvent things you never wanted touched. Describe the change. Name what stays. Stop there.

The Inputs, Field by Field

Here's the request shape on each::labs, using the real field names from the schema. Only two fields are required: prompt and videos.

videos: one base_video, up to four feature_video clips

The videos array must contain exactly one entry with "type": "base_video". That's the clip being edited. You can add feature_video entries next to it, up to five videos in total, to show the model a motion, a performance or a look you want carried into the edit. Every entry takes a url and an optional id. Files are MP4 or MOV, up to 300 MB, 3 to 30 seconds long, 24 to 60 fps, with each side between 700 and 4553 pixels and an aspect ratio between 2:5 and 5:2. One rule catches people: the combined duration of all videos (Video Character Elements included) has to land between 3 and 30 seconds, so a 25 second base clip leaves room for very little else.

images: frames and references

Up to 10 JPG or PNG images, each up to 30 MB and at least 300 by 300 pixels. Each image gets a type: first_frame, middle_frame and last_frame pin what the edited clip should look like at those moments, while refer_image is a general visual reference (the new jacket, the new product, the new location). If you send more than one middle_frame or more than one refer_image, each one needs an id.

elements and voices

elements takes up to 7 existing Kling Elements, each as an element_id from Kling's Element API plus your own id. Multi-image Elements and Video Character Elements are both supported, with at most 3 of the video kind. voices takes up to 3 existing custom or system voices as voice_id plus id. Across images, videos, Elements and voices, the total stays at 15 references or fewer, which matches the Omni Reference ceiling Kling announced.

The output settings

duration runs from 3 to 30 seconds (default 10). resolution is 720p or 1080p (default 1080p). aspect_ratio defaults to adaptive, with 21:9, 16:9, 9:16 and 1:1 available. The prompt holds up to 26,000 characters of English, and it's also where your negative descriptions go.

Name Your References, Then Use the Names

The id fields are the most underrated part of this endpoint. The schema describes them as reference identifiers used in the prompt, and they have to be unique across every input type in the task. That means you can write a prompt that says which asset does what, instead of hoping the model guesses which of your six attachments is the jacket.

Pick ids a human could read in a year: navy_jacket, new_bottle, lisbon_street. Not img1. Readable ids make the prompt read like a shot note. A request for a wardrobe swap might look like this:

{
  "prompt": "Replace the man's grey hoodie in source with the jacket shown in navy_jacket. Keep his face, hair, hands, the camera move and the timing exactly as in source. No logos on the jacket.",
  "videos": [
    { "type": "base_video", "id": "source", "url": "https://.../take-03.mp4" }
  ],
  "images": [
    { "type": "refer_image", "id": "navy_jacket", "url": "https://.../jacket-navy.png" }
  ],
  "duration": 8,
  "resolution": "1080p",
  "aspect_ratio": "adaptive"
}

Notice what the prompt doesn't do. It doesn't describe the room, the lighting or the man's age. The base video already carries all of that, and repeating it only creates chances for drift. Give the base video an id too, so you can say "as in source" and mean it.

A Kling Video Edit Workflow, Step by Step

Start by trimming the base clip to the moment you actually need. Every second you send counts against the 30 second budget, and a tighter clip gives the model less to reinterpret. Set duration to the length you want back.

Next, decide what kind of guidance the change needs. A single object or garment wants one refer_image. A change that has to look a certain way at a specific moment (a product label facing camera at the end, say) wants a last_frame. A recurring character you've already built in Kling wants an Element rather than a fresh photo, because an Element carries an identity you've already locked. If the edit involves movement you can't describe well in words, that's when a feature_video earns its slot.

Then write the prompt in three sentences: the change, the reference that defines it, and the list of things that must not move. The each::labs model page puts the same idea in fewer words: keep initial edits narrow, state the change first, and name the details that should stay the same.

Run it at 720p first. Once the edit reads right, rerun the winning prompt at 1080p. Processing is asynchronous, so build your integration to poll or wait rather than block. Finally, review the parts generative edits are most likely to disturb: faces, hands, lettering and product details. If one of them slipped, add it explicitly to the "keep" sentence and run again. Once a sequence of edits works, you can chain it with other steps in an each::labs flow so the next batch of takes goes through the same pipeline.

Five Kling 4 Edit Prompts to Steal

Each of these assumes the base video has "id": "source". Swap in your own ids.

Wardrobe swap

Change the woman's white blouse in source to the emerald silk dress shown in emerald_dress. Match the dress's drape and sheen to her movement. Keep her face, hair, jewellery, the background and the camera move unchanged.

Background replacement

Replace the office behind the speaker in source with the rooftop terrace shown in rooftop_ref, at golden hour. Relight the speaker so the warm light comes from frame left, matching the new setting. Keep his face, gestures, lip movement and framing exactly as in source.

Product swap

Replace the bottle the bartender picks up in source with the product shown in new_bottle. The label faces camera when he sets it down, as in end_frame. Keep his hands, the pour, the bar and the camera move unchanged. No extra text on the bottle.

Here end_frame is an image sent with "type": "last_frame", and new_bottle is a refer_image. Packaging is where you'll want to check the result closely, so keep that reference sharp and front-on.

Restyle

Transform source into a hand-inked watercolor animation with visible paper texture and soft bleeding edges. Keep the subject, the composition, the timing and the camera movement recognizable. No photoreal skin.

Add a character

Add the character from courier to source: he walks in from frame right at the start, stops beside the woman at the counter and hands her a parcel. He speaks one line in the voice courier_voice: "Signed for." Keep the woman, the shop and the camera move unchanged.

In this one, courier is a Video Character Element passed in elements, and courier_voice is an entry in voices. It's the edit that most needs its references, because the model has nothing in the base clip to borrow from.

Edit, Reference, Text or Image: Picking the Kling 4.0 Endpoint

The Kling 4.0 family on each::labs has four doors, and they share more plumbing than their names suggest.

Kling 4.0 Reference-to-Video takes the same inputs as Video Edit: images with the same four roles, videos, Elements, voices. The difference is what's required. Reference-to-Video only demands a prompt, and a base_video is allowed but optional. Video Edit demands one. So if you're building a new shot from references, start there. If there's a specific clip whose motion you're keeping, use Video Edit and make that contract explicit.

Kling 4.0 Text-to-Video is the bare version: a prompt plus duration, resolution and aspect ratio (16:9 by default, no adaptive option). It's what you use to make the first take. Kling 4.0 Image-to-Video requires 1 to 10 frames typed as first, middle or last frame, and it's the right call when you have a still you want to animate on a planned path.

A pattern that works well: generate the hero take with Text-to-Video or Image-to-Video, approve it, then spin out every variant (each colorway, each market, each season) with Video Edit. You lock motion once and edit everything else. For the bigger category view, browse video-to-video models on each::labs. If you're new to the release itself, our Kling 4.0 launch rundown covers the big picture, and the Kling 4.0 Prompting Guide covers writing shots from scratch.

Small Edits Win, Big Edits Wander

The temptation with a model this capable is to stack five changes into one request: new jacket, new street, new time of day, new style, and an extra character for good measure. Resist it. Every extra change is another place your base clip can drift. Make one change, approve it, and use the result as the base video for the next change. Five clean passes beat one ambitious mess.

Be specific about preservation, too. "Keep everything else the same" is a wish. "Keep his face, hands, the camera move and the timing" is an instruction. List the details the client cares about.

And use negative descriptions; the prompt field explicitly supports them. "No logos." "No extra people." "No photoreal skin." A short no-list closes doors the model would otherwise wander through. If you've worked with conversational editors, the rhythm will feel familiar: our Gemini Omni 1.1 Flash video editing guide makes the same case for one change per turn with Google's model.

Frequently Asked Questions

Is Kling 4 video editing the same as generating a new clip?

No, and that's the point of the endpoint. Video Edit requires a base_video, so the model works from your existing footage and its motion instead of inventing a shot from a description. You describe the difference, not the whole scene.

How long can the source video be in Kling video edit?

Between 3 and 30 seconds, and that budget is shared. The combined length of the base video, any feature videos and any Video Character Elements has to stay within 3 to 30 seconds, so trim the base clip before you add extras.

How do I tell Kling 4.0 which reference is which?

Give every input an id and use those ids in your prompt. Ids must be unique across images, videos, Elements and voices, and they're required when you send more than one middle_frame or refer_image. Readable names like new_bottle make the prompt clearer for you and for the model.

When should I use Reference-to-Video instead of Video Edit for AI video editing?

When there's no single clip you need to keep. Reference-to-Video accepts the same references but treats the base video as optional, which suits building a new shot. If a specific take's timing and camera move are the asset, Video Edit is the one built to start from that take.