How to edit a video with text prompts

You edit a video with a text prompt by sending the clip and one instruction to a video-to-video model, such as Kling O3 Edit, Luma Ray 2 Modify, Wan 2.7 Edit or Decart Lucy Edit. The model returns a new clip with the change applied and the original motion kept. At fal's list prices these models cost between $0.04 and $0.35 per second of footage. A five-second fix therefore costs between twenty cents and $1.75.

Getting a usable result depends less on the model than on three decisions: what kind of change you ask for, how long a clip you send, and whether the result replaces your shot or sits next to it. This guide covers all three. It also shows the NodeTool graph and the timeline flow that do it without a script.

Prices below are the per-second rates recorded in the provider catalogs NodeTool bills against, read in July 2026. You pay them on your own key, with no markup.

Quick answer

• Changing an object, outfit or prop inside a shot: Kling O3 Edit or Decart Lucy Edit Pro.
• Restyling the whole shot (live action to animation, day to night): Luma Ray 2 Modify, or Ray 2 Flash Modify while you iterate.
• Cheap drafts of an edit: Lucy Edit Fast at $0.04/s or Grok Imagine Edit at $0.05/s.
• A longer clip in one pass: Runway Aleph 2 on Replicate accepts 2 to 30 seconds, and Luma's Replicate endpoint accepts up to 30 seconds.
• Keeping a shot in the cut while you try edits: run the edit from the NodeTool timeline, where every result arrives as a candidate take.

Three kinds of edit, three kinds of model

A prompt edit is one of three jobs, and models are built for one of them.

1. Local edits. Swap the jacket, remove the cup, change the sign on the wall. Everything outside the change should stay pixel-close to the source. Kling's O1 and O3 edit endpoints, Decart Lucy Edit and Wan 2.7 Edit are built for this. Kling's prompt can refer to the input clip as @Video1 and to extra reference images as @Image1, which is how you say "replace the bag with this bag".
2. Restyles. Change the look of every frame and keep the motion. Luma describes Ray 2 Modify as "turning live-action into CG or stylized animation, to changing wardrobe, props". Lucy Restyle and the generic LTX video-to-video model in NodeTool's Video To Video node also sit here.
3. Structural edits. Make the shot longer, wider or a different shape. These are extend and reframe models: Veo 3.1 Extend, Grok Imagine Extend, Luma Ray 2 Reframe, Wan VACE outpainting and reframe. They do not change what is in the shot. They add more of it.

Sending a restyle prompt to a local-edit model, or the other way round, is the most common reason an edit "does nothing" or redraws the whole frame.

Edit models compared

Blank limits mean the node's schema states none. Runway Aleph is also reachable on kie as kie.video.GenerateAlephVideo. Wan 2.7 Edit runs on kie at $0.08/s and on AtlasCloud at $0.10/s, so the same edit costs a fifth less depending on which key you use.

Two things stand out in the table.

Length is the hidden constraint. Most local-edit models stop at 10 or 15 seconds. Cut your shot to the part that needs the change before you send it. That also cuts the bill, because every model here charges per second.

Draft cheap, finish once. Lucy Edit Fast at $0.04/s and Ray 2 Modify at $0.35/s differ by almost nine times. Use a cheap model to find the instruction that works, then run the final model once with that exact instruction.

Step 1: cut the clip to the change

Put a Trim node (nodetool.video.Trim) in front of the edit. Set the start and end to the range that needs editing, plus half a second either side. You will splice the edit back over the original, and the handles give you room for a clean cut.

If the source is a vertical or square crop of a wider shot, edit the wide shot first and reframe afterwards. Edit models see more context in a wider frame and invent less.

Step 2: write the instruction as a diff

An edit prompt describes the change, not the scene. The model already has the scene.

• Name what changes and what it becomes. "Replace the red umbrella with a clear plastic umbrella" beats "a woman with a clear umbrella in the rain".
• Name what must stay. "Keep her face, the street and the camera move unchanged." Local-edit models respect this. Restyle models will still repaint everything, which is the point of them.
• One change per pass. Two edits in one prompt usually get one done well and one done badly. Chain two edit nodes instead.
• Use references for anything branded. On Kling, attach the product photo and write "replace the bottle with @Image1". Words describe a product loosely. A reference image pins it.

Step 3: build the graph

The smallest editing graph is four nodes:

1. Video Input (nodetool.input.VideoInput) for the clip.
2. Trim to cut it to the range.
3. Video To Video (nodetool.video.VideoToVideo) with your instruction in prompt. Its model picker lists every video-to-video model your keys reach, and its strength dial (0 to 1, default 0.6) decides how far the result moves from the source.
4. Output (nodetool.output.Output) so the result is saved as an asset.

To use a provider-specific model with all its options, such as Kling's reference images or keep_audio, use that model's own node instead of the generic one. Search the node menu for the model name. The Video Restyle Studio template ships this graph with a Prompt node in front that expands a short style note into a full restyle brief.

Step 4: run it from the timeline instead

If the clip already sits in a cut, edit it there. Select a video clip in NodeTool's timeline, open AI Edit in the Inspector, and choose an operation: restyle, replace_object, remove_object, replace_range, extend or regenerate. NodeTool checks the operation against the selected model's capabilities before it spends anything, so asking a restyle-only model to remove an object fails up front instead of after the bill.

The result arrives as a candidate take on the clip. Your edit does not move: start time, duration, track and effects stay where they were. Choose Preview Candidate under History to audition it in place, then Use take if it belongs in the cut. If a generated take fails, the previous take stays active.

The agent can do the same from chat. "Make shot 3 darker and add rain" becomes a generative edit on that clip alone, and it waits for you to accept the take.

Step 5: finish the shot

Some edit models return a smaller frame than they were given. Grok Imagine Edit, for example, resizes its input to at most 854×480. Three steps close the gap:

• Upscale. Topaz video upscaling on fal costs $0.01/s, and the Topaz provider is also available directly. See video upscaling for the other options.
• Reframe. For a vertical version, the timeline's Adapt format makes 9:16, 1:1 and 4:5 sequences from the same media, or the Cut a Landscape Clip for Vertical template does a plain resize.
• Cut out. To put the edited subject on a new background, Bria and VEED background removal run as video-to-video nodes, and nodetool.video.ChromaKey handles green screen (template).

What a typical edit costs

Running all five passes of the first job on Kling O3 costs $3.50 instead of $1.50. The habit this table exists to break is drafting on the final model.

Checklist

1. Decide whether the change is local, a restyle, or structural.
2. Trim to the range, with half a second of handle on each side.
3. Write the change, not the scene, and say what must stay.
4. One change per pass. Chain nodes for more.
5. Attach a reference image for any product or logo.
6. Draft on a cheap model, finish once on the good one.
7. Audition the result as a take before it replaces anything.
8. Upscale and reframe last.