Kling 3.0 Omni Edit Video
One prompt, keep the sound, swap the world
- video
- reference_images
- VIDEO
Kling's edit line keeps the parts of your video you want and regenerates the parts you ask about - and the Omni tier is the current flagship of that idea. This node is the Kling 3.0 Omni video edit API, built into ComfyUI core as a partner/video node. No install, nothing on your GPU; the edit happens on Kling's servers through Comfy's proxy and bills against your Comfy account credits.
The shape is blessedly simple. video is what you're editing, model_name picks the 3.0 Omni edit model, and prompt is the instruction field - and note the tooltip: it can carry both positive and negative descriptions in the same box. "Turn the street into a rainy neon night, no people visible" is the kind of sentence it expects, mixing what to add with what to avoid without needing a separate negative field. Output length matches the input video, so what you feed in is what you get back, duration-wise.
The killer feature is keep_original_sound, defaulting to true. Most edit models throw away the original audio and force you to re-sync or re-design sound; Omni's whole pitch is that your dialogue and ambience survive the edit. Turn it off only if you want the model to start fresh. For any edit that involves a character talking, leaving this on is usually the right call - it's the difference between an edit and a re-shoot.
The optional extras punch above their weight: reference_images accepts up to 4 additional images to guide the edit - style references, a face to composite in, a prop to match - and resolution is there if you want to control output size. seed is the standard non-deterministic re-run switch.
Output is a single VIDEO. The workflow that makes this shine is the "change the environment, keep the performance" edit: an actor's take, a product shot, a location plate - prompt the new setting, keep the sound, and you have a usable alternate take without reshooting or regenerating the subject's motion. It's the same philosophy as Wan 2.7's video edit but with sound preservation front and center, which is a genuine differentiator for dialogue-driven work.
Honest caveats, all standard for the family: the input video and reference images are uploaded to Kling's servers, moderation runs on the vendor side (a rejection there is the filter, not your graph), and you should draft cheap - short clips, lower resolution - before committing a long edit at max settings. The price badge updates as you change resolution, and per-render billing means iteration strategy matters more than it does for a local node.
Omni edit is a newer arrival in Kling's core lineup, following Kling's Omni model line through 2025–2026. If your video is good but the world around it is wrong, this is the node that fixes the world and leaves the performance alone.
Inputs (7)
| Name | Type | Default | Description |
|---|---|---|---|
| model_name | COMBO | 2 options: kling-v3-omni, kling-video-o1 | |
| prompt | STRING | A text prompt describing the video content. This can include both positive and negative descriptions. | |
| video | VIDEO | Video for editing. The output video length will be the same. | |
| keep_original_sound | BOOLEAN | true | — |
| reference_imagesopt | IMAGE | Up to 4 additional reference images. | |
| resolutionopt | COMBO | 2 options: 1080p, 720p | |
| seedopt | INT | 00–2147483647 | Seed controls whether the node should re-run; results are non-deterministic regardless of seed. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| VIDEO | VIDEO | — |