Kling 3.0 Omni API

Kling 3.0 Omni is the reference-driven half of Kuaishou's Kling 3.0 release, and the model in the family built for keeping things the same. Hand it a few photos of a character, a product or a location, or a short video of a person talking, and it carries their look, and even their voice, into a new scene. It writes the dialogue track and sound effects in the same pass as the picture, can cut between up to six shots inside one clip, and can also edit footage you already have.

Kuaishou announced it on 5 February 2026 alongside Kling Video 3.0, the prompt-led sibling, and two matching image models. Since then the Omni model has gained native 4K generation in April and, on 17 June, an upgrade aimed squarely at editing: edits that stay closer to the source, edit lengths from 3 to 15 seconds, and 4K input and output.

That combination makes Omni the Kling model to reach for when a project has continuity requirements: a recurring character in a series, a product that must look identical in every ad, or a shot that needs changing rather than replacing. On Unicorn API it is a single model, kling-3.0-omni/video, that covers text-to-video, image-to-video, reference-to-video and video transformation.

Kuaishou, Kling and the road from Kling O1

Kling is the AI video business of Kuaishou Technology, the Beijing short-video company. It launched in June 2024, and by the 3.0 announcement Kuaishou said it served more than 60 million creators, had produced over 600 million videos and worked with more than 30,000 enterprise clients. Kuaishou also said Kling's annualised revenue run rate reached 240 million US dollars in December 2025, which makes it one of the few AI video products with a large paying business behind it.

The Kling line moved quickly: 1.0 in mid-2024, 1.6 at the end of that year, 2.0 and 2.1 in spring 2025, then 2.5 Turbo. In early December 2025 Kling held an "Omni Launch Week" that produced two branches. Kling 2.6 brought native audio to the classic text- and image-to-video models. Kling O1, short for Omni One, was presented as the first unified multimodal video model: one engine that accepted text, images, video and subjects together and handled generation, reference-based creation, start and end frames, inpainting, restyling and shot extension.

O1 had clear gaps. It generated 3 to 10 seconds, with no native audio and no multi-shot. Kling 3.0 Omni is the direct answer to those gaps, merging the Omni engine's reference and editing skills with the audio and storytelling features of the main Kling line.

Like every Kling model, 3.0 Omni is closed. Kuaishou has not released weights or a parameter count, and the model is available only through Kling's own apps, the Kling API and partners such as Unicorn. At launch it went to Kling's Ultra subscribers first, with wider availability following.

What changed in Kling 3.0 Omni

Sound in the same pass

O1 produced silent video. 3.0 Omni generates dialogue, sound effects and ambience together with the picture, with lip sync for speaking characters. Speech is supported in Chinese, English, Japanese, Korean and Spanish, with American, British and Indian English accents and several Chinese dialects, and Kuaishou says characters in one scene can each speak a different language.

Up to six shots in one clip

The headline feature of the 3.0 family is multi-shot generation. Instead of one continuous take, a single clip can contain up to six shots, and for each one you can set duration, shot size, perspective, what happens and how the camera moves. If you would rather not plan the cuts, the model can split a prompt into shots on its own. Kuaishou pitches this as moving from clip generation to scene generation, with patterns such as shot-reverse-shot dialogue and cross-cutting built in.

Longer clips

Length goes from O1's 10-second ceiling to 15 seconds, selectable in whole seconds from 3 upward. That is enough for a short dialogue exchange or a complete product spot.

Characters with a voice

Omni's most distinctive addition is the video element. Upload a 3 to 8 second clip of a character and the model extracts both appearance and voice, then reuses them in new scenes. Kling calls the voice side Element Voice Control: the timbre stays attached to the character rather than being reinvented each generation. Elements can also be built from two to four photos taken from different angles.

More references at once

A generation can draw on up to seven reference images or elements, or up to four when a reference video is also supplied. A reference video of 3 to 10 seconds can serve either as the footage to edit or as a guide for style and camera movement.

4K and the June editing upgrade

On 23 April Kling added native 4K generation to the 3.0 video series, rendering 3840 by 2160 directly rather than upscaling. The 17 June upgrade then turned to editing: Kling says Omni now understands input videos and images significantly better, so edits stay faithful to the source, and it accepts and returns edits of 3 to 15 seconds in up to 4K.

How the Omni model works

Kuaishou has not published a technical report specific to 3.0 Omni, but it described the Omni architecture in its Kling-Omni technical report in December 2025. There, a multimodal language model acts as a prompt enhancer that interprets mixed instructions; an Omni-Generator processes visual and text tokens in a shared embedding space, which is what lets one model treat "make this", "keep this face" and "change this shot" as the same kind of request; and a cascaded super-resolution stage refines detail. Training combined large-scale text-video pre-training, fine-tuning on reference and editing tasks, preference optimisation for motion and visual integrity, and distillation that cut sampling from 150 model evaluations to 10. Kling's earlier models are diffusion transformers with an in-house 3D VAE. How much of this changed for 3.0, and how large the model is, Kuaishou has not said.

Kling 3.0 Omni specs on Unicorn

Pick the workflow with mode: t2v, i2v, r2v or transform. Turn on sound with with_audio, let the model plan cuts with prefer_multi_shots, or write your own shots in multi_prompt with customize_multi_shots. When you transform a video without adding images, the output follows the length of the source clip.

Kling 3.0 or Kling 3.0 Omni

Kuaishou frames the split simply. Kling 3.0 is prompt-led: text, a start frame or start and end frames, with the strongest tools for writing a scene from words. Omni is reference-led: it adds multi-image references, video elements and voice control, and it can edit existing footage. Both share the 15-second limit, native audio and multi-shot. If the brief is "invent this", start with Kling 3.0; if it is "use these" or "change that", use Omni.

Where Kling 3.0 Omni shines

Recurring characters. Series, short dramas and mascots need the same face and the same voice every episode. Video elements carry both from one short clip.

Products and brands. With seven reference slots, a product shot, a packaging close-up and a location can all be locked in, which suits e-commerce and ad variants. Kling also says text on products stays legible across shots.

Dialogue scenes. Native multilingual speech plus multi-shot means a two-person exchange with cuts can come out of one generation, sound included.

Editing footage. Transform mode changes an existing clip, its setting, style or details, instead of regenerating from scratch, and since June it handles up to 15 seconds and 4K.

Testers' verdicts. Reviewers singled out the multi-shot structure as a major differentiator, with one calling the result "closer to a scene" than a clip, and Curious Refuge called Kling 3.0's image-to-video "easily the highest-scoring AI video model we've reviewed to date". In blind voting on the Artificial Analysis text-to-video arena, Omni at 1080p edges standard Kling 3.0 at the same resolution.

Limits to know before you build on it

Fifteen seconds is the ceiling. Anything longer is assembled from several generations, and continuity between them is your job, helped by elements.

No generated audio with a reference video. When a video is part of the input, native sound generation is off; plan to add audio separately or use image references instead.

Lip sync and cloning are uneven. Curious Refuge found lip sync "doesn't always hit the mark" and called character cloning closer to "an R&D tool" than a production asset because of likeness drift. Some testers report voices overlapping when more than two characters speak.

Big edits can break. Heavy scene changes, crowds and effects such as rain produced artefacts and distorted characters in tests. Omni is strongest when an edit is targeted.

Inputs decide quality. Clean, front-facing, well-lit references on uncluttered backgrounds work best; busy reference images confuse the model.

Three aspect ratios. 16:9, 9:16 and 1:1 only.

Prompting tips for Kling 3.0 Omni

Write like a director. Name the shot size and camera move, describe the action in order and put spoken lines in quotation marks with the speaker identified. When you use references, say which image is which in plain words ("the woman in the first reference image", "the street in the second") and give each a role. Keep references consistent in style: a photo character in a painted location tends to fight.

A multi-shot dialogue prompt with sound:

Shot 1: wide shot of a small bakery at dawn, the baker slides a tray of bread onto the counter. Shot 2: close-up of a young customer who smiles and says, "Is it always this early?" Shot 3: the baker laughs and replies, "Only for the good bread." Warm light, quiet street ambience.

A reference-to-video prompt:

The cyclist from the first reference image rides through the mountain village from the second reference image at golden hour. Tracking shot from the side, then a low-angle shot as she crests the hill. Keep her red jacket and helmet exactly as in the reference.

A transform prompt for an existing clip:

Keep the camera move and the actor's performance. Change the season to deep winter: snow on the ground and roofs, breath visible in the air, overcast light.

Draft at 720p and short durations while you adjust blocking and timing, then render the keeper longer and at 1080p or 4K.

Kling 3.0 Omni against other video models

Kling 3.0 shares the same 15-second limit, native audio and multi-shot, and is the better starting point when you work from text or a start and end frame.

Seedance 2.0 from ByteDance is a strong all-rounder, worth testing side by side for prompt-led shots.

Wan 2.7 Reference to Video and HappyHorse Reference To Video are the other reference-driven options; compare them on the same character sheet when consistency is the whole job.

Kling 2.6 remains a simpler Kling with native audio for single-shot clips of 5 or 10 seconds that need no references.

The rule of thumb: choose Omni when your inputs matter as much as your prompt, meaning a face, a voice, a product or a piece of footage that must survive into the result.

Getting started

Open Kling 3.0 Omni in the studio, choose create, animate, references or transform, and start at 720p. For code, create a key on the API keys page and submit a job to /v1/jobs with model set to kling-3.0-omni/video, a prompt, a mode, and the inputs that mode needs: image and optional end_image for animation, reference_images, reference_videos or elements for references and transformation. The API tab on this page has full request examples, and the docs explain how jobs work.

Verdict

Kling 3.0 Omni is where Kuaishou's two video branches meet: the Omni engine's references and editing, and the 3.0 line's sound, multi-shot storytelling and 15-second clips. Video elements that carry a voice, seven reference slots, 4K and a June editing upgrade make it one of the most controllable video models available. Lip sync, cloning and large edits still need checking before anything ships, and audio cannot accompany a reference video. But for recurring characters, branded products and changes to footage you already have, Omni is the Kling model to start with.