# How to Edit a Video With Text Prompts: Kling O3 Edit, Gemini Omni, Grok and Luma Tested

Canonical page: https://fuser.studio/articles/edit-video-with-text-prompts

How to change an existing video with a text instruction: input limits, prompt structure, and Kling O3 Edit vs Gemini Omni, Grok and Luma on the same clip.

[All guides](https://fuser.studio/articles) · [Best AI video editing models](https://fuser.studio/articles/best-ai-video-editing-models)

**Quick answer:** To change a video you already have with a text instruction, connect the clip to an editing model, write one change plus a sentence naming what must stay, and compare the output against the source frame by frame. In Fuser the editors that work this way are [Kling O3 Edit](https://fuser.studio/models/kling-o3-edit), [Gemini Omni Video](https://fuser.studio/models/gemini-omni-video), [Grok Imagine Edit](https://fuser.studio/models/grok-imagine-edit), [Luma Ray](https://fuser.studio/models/luma-ray) in Video Edit mode and [Runway Aleph](https://fuser.studio/models/runway-aleph); the [LTX 2.5](https://fuser.studio/models/ltx-2-5) node retakes only a section of a clip. We gave the same five-second take and the same two instructions to Kling O3 Edit, Gemini Omni 1.1 Flash, Grok Imagine Edit and Luma Ray 3.2, one run each. Kling O3 Edit Standard produced the most convincing night re-light while keeping the performer, set, timing and sound, and a watercolour restyle that held the scene with a few cosmetic additions. Gemini Omni 1.1 Flash was the most literal on both.

## 1. Decide what kind of change you need

A text-instruction edit reworks the whole clip you give it: the lighting, the style, a costume, the weather, an object's colour. Kling's guide for its Omni editing describes exactly this kind of request, with examples like "remove bystanders", "change daytime to dusk" or "replace the main character's outfit" ([Kling VIDEO O1 guide](https://kling.ai/quickstart/klingai-video-o1-user-guide)), and says that video editing in VIDEO 3.0 Omni, the model Kling O3 Edit runs, and other features "function the same as in O1" ([Kling VIDEO 3.0 Omni guide](https://kling.ai/quickstart/klingai-video-3-omni-model-user-guide)).

Two neighbouring jobs need a different tool:

- **Only a few seconds are wrong**, such as one line of dialogue or the ending. Retake that window instead of editing the whole take. Connect the clip to an LTX 2.5 node and it switches to retake, which runs the LTX-2.3 retake model and replaces the picture, the sound or both in the window you choose ([LTX retake API](https://docs.ltx.io/api-documentation/api-reference/async-video-generation/submit-retake)). In Fuser the window is 6 to 20 seconds long and starts on a whole second, so retake suits clips longer than the 5-second one we tested here.
- **You need a new shot that matches this one**, such as the next shot. Switch Kling O3 Edit from Edit Video to Reference Video mode; it generates a new 3 to 15 second shot guided by the clip's motion and camera. Our [Kling O3 Edit guide](https://fuser.studio/articles/kling-o3-edit-guide) shows a Reference run. If the shot just needs to be longer, see [how to extend AI video length](https://fuser.studio/articles/extend-ai-video-length).

## 2. Check the source clip against the editor's limits

Most failed edits fail before the model runs, on the input. These are the limits on the Fuser nodes:

- **Kling O3 Edit:** 3 to 15 seconds, MP4 or MOV, at least 720 pixels on both sides (up to 3,840), 24 to 60 fps, up to 200 MB. Kling's own guides list 3 to 10 seconds and up to 2K for its app ([Kling VIDEO 3.0 Omni guide](https://kling.ai/quickstart/klingai-video-3-omni-model-user-guide)), so treat clips over 10 seconds with some caution. In Edit mode the output follows the input's length.
- **Gemini Omni Video:** up to 10 seconds; the node refuses a longer clip before it runs. Google gives the same 10-second limit for edits ([Gemini API: Omni](https://ai.google.dev/gemini-api/docs/omni)).
- **Grok Imagine Edit:** the node resizes the clip to a maximum area of 854 × 480 pixels and uses the first 8 seconds. xAI says the output takes the input's duration and aspect ratio, capped at 720p ([xAI video editing docs](https://docs.x.ai/developers/model-capabilities/video/editing)).
- **Luma Ray, Video Edit mode:** returns 5 or 10 seconds at 540p, 720p or 1080p.
- **Runway Aleph:** Runway's API accepts 2 to 30 seconds for Aleph 2.0, the longest input here ([Runway API changelog](https://docs.dev.runwayml.com/api-details/api_changelog/)).

The 720-pixel minimum is the one that catches people with Kling O3 Edit: a 480p clip, or a vertical clip narrower than 720 pixels, is rejected. Export at 720p or higher, or run the clip through an upscaler node such as [SeedVR](https://fuser.studio/models/seedvr-upscale) or [Topaz](https://fuser.studio/models/topaz-upscale) first. Our test clip, a 5-second take of a ceramicist at a pottery wheel who looks up and says "Almost there" (generated with Veo 3.1 for [our model comparison](https://fuser.studio/articles/seedance-vs-kling-vs-veo) and trimmed), is 1280 × 720 at 24 fps with sound, so it passed every editor without re-encoding.

## 3. Write one change, then say what stays

Both of our instructions use one structure: name the single change, then add a clause listing what must not change. Kling O3 Edit, Gemini Omni and Grok Imagine Edit kept the elements we listed; Luma Ray 3.2, which asks for a different prompt style (below), did not.

- Re-light: "Re-light the scene as late evening: outside the window it is dark blue night, and the workshop is lit only by the warm desk lamp above the wheel. Keep the ceramicist, her movements, the wheel, the vase and the camera exactly as they are."
- Restyle: "Restyle the whole clip as a hand-painted watercolour animation with soft paper texture and loose ink outlines. Keep the ceramicist, her movements, the wheel, the vase and the camera exactly as they are."

Each vendor adds its own grammar on top:

- **Kling O3 Edit** refers to the connected clip as @Video1 and to reference images as @Image1 to @Image4, and Kling's own templates are one line each, such as "Remove [describe content to remove] from [@Video]" and "Change the [item] in the [@Video] to [color]" ([Kling VIDEO O1 guide](https://kling.ai/quickstart/klingai-video-o1-user-guide)). The prompt can run to 2,500 characters, but the templates show it doesn't need to.
- **Gemini Omni:** Google says "Simple prompts work best for video editing. Overly descriptive prompts can lead to unintended changes", and suggests ending with "Keep everything else the same" ([Gemini API: Omni](https://ai.google.dev/gemini-api/docs/omni)).
- **Luma Ray 3.2** wants the opposite of an instruction: "Your prompt should describe the target end state. Do not write commands", and no negations such as "no" or "without" ([Luma: Ray 3.2 video to video](https://lumalabs.ai/learning-center/articles/ray-3-2-video-to-video)). Write "a pottery workshop late at night, lit by a desk lamp" rather than "re-light the scene".
- **Runway Aleph:** the Fuser node's help asks for an action verb for the scope of the change, such as add, remove, change, replace, re-light or re-style, followed by the result you want. Prompts are capped at 1,000 characters.

## 4. Run the same instruction through several editors

Editors differ most in what they leave alone, and you only see that by comparing them on your own footage. On a Fuser canvas, connect one video node and one text node to several editing nodes and run them together. We gave the clip above to each editor with the instructions exactly as written in step 3. Kling O3 Edit ran on 2 October 2026 on Standard with Keep Audio on, one run per instruction; the Gemini Omni 1.1 Flash, Grok Imagine Edit and Luma Ray 3.2 outputs are the 720p runs from our [video editing model test](https://fuser.studio/articles/best-ai-video-editing-models) on 28 September 2026, one run each, and the Kling O3 Edit Pro re-light comes from our [Kling O3 Edit guide](https://fuser.studio/articles/kling-o3-edit-guide). The Gemini edits were generated with the same model version Fuser's node calls.

### Instruction 1: re-light the take as night

![The original daytime pottery take beside night re-lights from Kling O3 Edit Standard and Pro, Gemini Omni 1.1 Flash, Grok Imagine Edit and Luma Ray 3.2.](https://statics.fuser.studio/cms/646b4762-e26d-4b2d-9307-c8b577f9de2f)

_Instruction 1, same 5 s 720p source and prompt, one run per model. Kling O3 Edit: 2 October 2026 (Pro scaled down from 1080p). Others: 28 September 2026, 720p; Luma is its shared-prompt run. Plays the source soundtrack; the Grok tile holds its last frame for the final 0.3 s._

- **Kling O3 Edit, Standard** delivered the night the prompt asked for: a deep blue window, the desk lamp as the main light on the clay and her hands, and the rest of the workshop in darkness. The ceramicist, her apron and clip-on microphone, the shelves, the vase's shape, her look up at the camera and her smile all land where they do in the source.
- **Kling O3 Edit, Pro** gave a very similar result. The visible difference was resolution: Pro returned 1920 × 1080 from the 720p source, Standard 1280 × 720.
- **Gemini Omni 1.1 Flash** kept every detail of the shot, with a dark blue window and the lamp on, but the room stayed brighter than "lit only by the lamp".
- **Grok Imagine Edit** kept the shot too, but read the light as a warm dusk rather than dark blue night.
- **Luma Ray 3.2** made a dramatic night but rebuilt the scene: a large paned window replaced the back wall, the shelves went, and the ceramicist's face and hair changed. A second Luma run with an end-state prompt held the set but changed her into a different person ([details in our editing model test](https://fuser.studio/articles/best-ai-video-editing-models)).

### Instruction 2: restyle the clip as watercolour

![The original pottery take beside watercolour restyles from Kling O3 Edit, Gemini Omni 1.1 Flash, Grok Imagine Edit and Luma Ray 3.2.](https://statics.fuser.studio/cms/2fb54380-3ca0-47fc-9b6c-7e75ddc95a40)

_Instruction 2, same source and prompt, one run per model. Kling O3 Edit Standard: 2 October 2026. Gemini Omni 1.1 Flash, Grok Imagine Edit and Luma Ray 3.2: 28 September 2026, 720p. The grid plays the source soundtrack; the Grok tile holds its last frame for the final 0.3 s._

- **Kling O3 Edit, Standard** produced a clear watercolour look, with washes, ink outlines and a paper ground, and kept the composition: the same woman and movements, the apron straps and microphone, the shelves, window and lamp, and a vase that stays a vase. It also took some liberties: red lipstick and blushed cheeks, a few extra pots on the shelves, and the clay smear on her forearm turned into a dark mark that reads like a tattoo.
- **Gemini Omni 1.1 Flash** gave the most literal restyle: watercolour and ink over the same woman, apron, shelves, brush pots and lamp, with nothing added.
- **Grok Imagine Edit** produced a convincing watercolour and kept the set, but gave the vase two handles, turning it into a jug.
- **Luma Ray 3.2** made an inked illustration rather than a watercolour and redesigned the ceramicist, with a headscarf and a blue shirt.

Two instructions with one run each is a small sample, so read these as how each model behaved on this clip, not a ranking. On it, Kling O3 Edit and Gemini Omni 1.1 Flash were the two that changed the look while keeping the shot; Kling committed harder to the night, and Gemini changed less in the restyle.

## 5. Check the output against the source

Scrub the edited clip next to the original before you use it. The problems our runs produced are the ones to look for:

- **Objects that change shape:** Grok's vase grew handles in the restyle.
- **People who change identity:** both Luma re-lights changed the performer. If the face matters, describe the person in the prompt for models that rebuild the scene.
- **Small additions:** Kling's restyle added make-up and a forearm mark nobody asked for.
- **Length:** Grok's outputs ran 4.71 seconds against the 5-second source. Kling's ran 5.04 seconds, and Gemini's and Luma's the full 5.
- **Sound:** with Keep Audio on, both Kling outputs carried the source soundtrack at the same loudness as the original, so the dialogue still matches the unchanged mouth movement. Google notes that Gemini Omni doesn't support voice editing ([Gemini API: Omni](https://ai.google.dev/gemini-api/docs/omni)), so change dialogue with an LTX retake instead.

## 6. Compare the cost before you pick a tier

By the credit cost Fuser's nodes calculate for a 5-second 720p source like ours, Grok Imagine Edit is the cheapest editor here. A Kling O3 Edit Standard edit costs about the same as a 720p Gemini Omni edit, about one and a half times a Grok edit, a little over half a 720p Luma Ray edit and under half a Runway Aleph edit. Pro costs a third more than Standard. Kling O3 Edit is charged per whole second of input, so a 10-second clip costs twice as much as a 5-second one. A practical routine is to find the instruction on Standard, then rerun the take you keep on Pro if you need the 1080p output.

## 7. Finish the clip in the same graph

The edited clip is an ordinary video output, so it can feed the next node. Upscale it with [Topaz](https://fuser.studio/models/topaz-upscale) or [SeedVR](https://fuser.studio/models/seedvr-upscale), add sound with [MMAudio](https://fuser.studio/models/mmaudio) if you edited with Keep Audio off, caption it from its soundtrack with [Auto Caption](https://fuser.studio/models/auto-caption) to check the dialogue survived, or send it to an LTX 2.5 node to retake a section. For more on writing Kling instructions see the [Kling O3 Edit guide](https://fuser.studio/articles/kling-o3-edit-guide); for the other editors in depth, the [Luma Ray 3.2 guide](https://fuser.studio/articles/luma-ray-3-2-guide), the [Runway Aleph 2.0 guide](https://fuser.studio/articles/runway-aleph-2-guide) and the [LTX-2.5 guide](https://fuser.studio/articles/ltx-2-5-guide).

## Which editor to reach for.

Test notes from our two instructions on one 5-second clip, one run each; limits from the Fuser nodes and vendor docs.

### Text-instruction video editors in Fuser

| Editor | On our two instructions | Source limits in Fuser |
| --- | --- | --- |
| Kling O3 Edit (Standard) | Strongest night re-light that kept the shot; restyle kept the scene but added make-up and a forearm mark. Sound kept with Keep Audio. | 3 to 15 s, at least 720 px on both sides, 24 to 60 fps, up to 200 MB; up to 4 reference images |
| Gemini Omni 1.1 Flash | Most literal on both: kept every detail; the re-light stayed a little bright. | Up to 10 s; 360p to 4K out, above 720p upscaled |
| Grok Imagine Edit | Kept the shot; dusk rather than night; the restyle gave the vase handles. Cheapest per clip. | First 8 s, resized to a maximum area of 854 × 480; 480p or 720p out |
| Luma Ray 3.2 (Video Edit) | Boldest change; rebuilt the set and changed the performer in both. | 5 or 10 s out at 540p, 720p or 1080p; Adhere, Flex or Reimagine strength |
| Runway Aleph 2.0 | Not run in our test. | 2 to 30 s input; one reference image pinned to the first frame |
| LTX 2.5 node, retake (LTX-2.3) | Not used here: the shortest retake window is longer than our clip. | Replaces picture, sound or both in a 6 to 20 s window |

## Questions, answered.

### Can AI edit an existing video from a text prompt?

Yes. Video-to-video editors such as Kling O3 Edit, Gemini Omni, Grok Imagine Edit, Luma Ray 3.2 and Runway Aleph take a clip and a written instruction and return the clip with that change. In our test, Kling O3 Edit and Gemini Omni 1.1 Flash re-lit and restyled a 5-second take while keeping the performer, set and timing.

### How do I write a prompt for AI video editing?

Ask for one change, then name what must stay, for example "Keep the ceramicist, her movements, the wheel, the vase and the camera exactly as they are." For Kling O3 Edit, refer to the clip as @Video1. For Luma Ray 3.2, describe the finished scene instead of giving a command.

### Why does Kling O3 Edit reject my video?

Check the limits: 3 to 15 seconds, at least 720 pixels on both sides, 24 to 60 fps, MP4 or MOV, up to 200 MB. A 480p clip fails the size check; export at 720p or upscale it first.

### Does editing a video with AI keep the original audio?

It can. With Keep Audio on, Kling O3 Edit returned our source soundtrack at the same loudness, and the other editors we tested also kept it. Google says Gemini Omni does not support voice editing, so use an LTX retake to change dialogue.

### Can I change only part of a video?

Yes. Connect the clip to Fuser’s LTX 2.5 node and it retakes a window you choose, 6 to 20 seconds long, replacing the picture, the sound or both and keeping the rest.

### Which AI video editor is cheapest?

Of the editors here, Grok Imagine Edit costs the least per clip. A Kling O3 Edit Standard edit costs about the same as a 720p Gemini Omni edit and under half a Runway Aleph edit; Kling O3 Edit Pro costs a third more than Standard.

## Edit one take with every model.

Wire a clip and one instruction into Kling O3 Edit, Gemini Omni, Grok and Luma on one canvas, and keep the version that changed only what you asked.

[Open Kling O3 Edit in Fuser](https://fuser.studio/models/kling-o3-edit) · [Explore all guides](https://fuser.studio/articles)

## More articles

- [Best AI Video Editing Models: Same Clip, Same Edits](https://fuser.studio/articles/best-ai-video-editing-models.md)
- [Grok Imagine Video 1.5 Guide: Text, Image, Edit and Extend](https://fuser.studio/articles/grok-imagine-video-guide.md)
- [Kling O3 vs Kling 3.0: Which to Use, Tested on the Same Inputs](https://fuser.studio/articles/kling-o3-vs-kling-3.md)