How to re-shoot a clip from a written instruction

This is not trimming and it is not compositing. You upload a clip, write what should be different about it, and the model renders the shot again with that change made — the same action, in a different coat, in a different room, in a different grade. It is closer to a re-shoot than to an edit, which is why it is worth being precise about what it can and cannot reach.

What a prompt edit is, and is not

AI edit takes the whole clip as its subject. It reads the motion and the composition out of your footage and re-renders the shot with your instruction applied. Because it is a re-render rather than a mask-and-patch, the change can be structural — a different setting, a different outfit, a different time of day — in a way a compositing tool could not manage.

The same fact is its limitation. Every frame is generated, so nothing survives untouched. If you need the original pixels preserved outside the change, this is the wrong tool.

Which plan: Starter and up, paid in ✦ gold credits. Kling O1 Edit is a hosted model billed per call by its provider — the green wallet cannot buy it, and signing up grants no gold.

The envelope

Kling O1 Edit

Property	Value
Source clip	required, 3 to 10 seconds
Reference images	up to four, optional
Output length	the source clip's length
Resolution	720p
Wallet	✦ gold, priced per second of output
Plan	Starter and up

Ten seconds is a hard ceiling, and it is the first thing to check against your footage. A longer clip has to be cut down before it can be edited at all — which, on this platform, means cutting it somewhere else first, because the in-app timeline is not available below Ultra.

What it costs: Priced per second of output in gold, so a ten-second edit costs roughly three times a three-second one. Test the instruction on the shortest usable trim before spending on the full clip. The quote is on the Generate button before you press it.

Running an edit

1. Open the AI edit section. Go to /video and pick AI edit from the rail.
2. Bring in a clip of 3 to 10 seconds. Upload it or pick it from your asset library. Anything longer is refused — trim first.
3. Write one change. Name what should be different and leave everything else unmentioned. "Change the jacket to red leather" works. A paragraph listing six changes gets you an average of six half-changes.
4. Attach references if the change has a specific look. Up to four images. This is how you pin an exact product, garment or face rather than describing it and hoping.
5. Check the quote and generate. Cost is per second of the source clip's length and shown before you press.
6. Compare, then iterate on the prompt. The result is a new asset alongside the original, so you can keep both. Change one thing in the instruction per attempt — it is the only way to learn what the model heard.

Open AI edit — Opens the AI edit section. Upload a clip and write the change; nothing runs until you press Generate.

Writing an instruction the model can act on

• Name the target and the change. "The wall behind her becomes exposed brick" gives the model both. "Make it grittier" gives it neither.
• One change per render. Compound instructions get averaged. Chain two edits instead — the output of one is a valid input to the next, at the cost of a second render.
• Do not re-describe what should stay. Everything you do not mention is left to the source clip, which is the behaviour you want.
• Use references for identity. A described product is a similar product. An attached photograph of it is that product.
• Expect the whole frame to move. Because every frame is re-rendered, fine detail outside the change will shift slightly. Choose the shot accordingly.

What it is actually good at

The instruction format rewards a particular kind of job: one where the action in the footage is right and the world around it is wrong. Those are the cases worth spending gold on.

• Restaging. The same performance in a different room, on a different street, at a different time of day.
• Wardrobe and product swaps. A different jacket, a different can in the hand, a different label — this is where the reference images earn their place.
• Grade and era. Turning a flat clip into something that reads as film, or as security footage, or as the nineties.
• Weather and light. Adding rain, moving the sun, changing an interior from daylight to lamplight.
• Style. Pushing a live-action clip toward illustration or a rendered look while keeping the motion.

What it is not good at is anything where the original pixels have to survive: matching a real product exactly, keeping small on-screen text legible, or preserving a face at the fidelity a client would check. Every frame is generated, so nothing is preserved by default — an attached reference improves the odds enormously, but it is an argument to the model, not a guarantee.

Using the four reference slots

Four images ride along with the instruction, and they are the difference between a described change and a specific one. A prompt that says "a red leather jacket" gets a red leather jacket; the same prompt with a photograph of the jacket attached gets that jacket, with its collar and its zips.

Use them for identity — a face, a product, a logo, a specific location — and spend the prompt on the relationship instead: what the referenced thing is doing in the shot, and where it sits. Four is the ceiling, and in practice two clear references beat four muddy ones, because every additional image is another constraint the model has to satisfy at once.

The Timeline editor is a different thing, and it is Ultra only

It is worth stating plainly, because the two get confused. AI edit is generative: it re-renders a clip from an instruction. The Timeline editor is a conventional editor — multi-segment trim, split by scene, crop, rotate, speed, fades, joining clips from several sources, GIF export — and it does not call a model or spend credits at all.

Timeline is an Ultra-only section, and it is not merely locked below that plan, it is not shown at all. If you are on Starter, Creator or Pro, plan on trimming and assembling in whatever editor you already use, and treat this platform as the place the generated shots come from.

Three things people mean by "edit the video"

You want to	Use	Plan
Change what is in the shot	AI edit — Kling O1 Edit	Starter and up, gold
Make a character perform a movement	Motion	Starter and up, gold
Trim, split, speed, crop, join	Timeline	Ultra only
Cut a long source into shorts	Analyze, then export the picks	Starter and up, gold

The last row is the one people miss. If the job is really "take this hour of footage and get the good ten seconds out of it", that is not an edit at all — it is a search, and Analyze does it with a timestamped index and a plain-language query, with the export costing nothing.

Watch out: Video Upscale, Watermark Remove and Subtitle Remove are also Ultra-only sections and are likewise invisible below that plan. Do not design a workflow around them from a lower tier.

What plan do I need?

Starter. Kling O1 Edit is a hosted model billed in gold credits, and a free account is granted none. There is no own-GPU version of this lane.

How long a clip can I edit?

Three to ten seconds. Longer footage has to be cut down first, and the in-app timeline that would do that is Ultra only, so use your own editor for the trim.

Does the original clip survive?

The original asset stays in your library untouched — the edit lands as a new asset beside it. Within the new file, though, every frame has been re-rendered.

Can I change a person's clothes but keep their face?

That is exactly what the reference images are for: attach a photograph of the face alongside the garment instruction. Identity holds much better with a reference than with a description.

Can I use it to remove a watermark or subtitles from a clip?

No. Those are separate, purpose-built lanes and both are Ultra-only sections, invisible below that plan. Their output is also capped at a small short side and is never upscaled.