Kuaishou video model · official name Kling 3.0 Omni · kling-v3-omni

Kling O3 Video Generator

Run the reference-driven tier of Kuaishou’s Kling 3.0 generation, the only model in the six-model capability map whose rows mark video reference and multi-image input supported, billed by input combination.

Try Kling O3
01

Create with Kling O3

Pick text, a single image, or a reference image, set duration, resolution, and aspect ratio, then read the point estimate before you submit.

LIVE
2s
2s15s

Results

Creation history

Generated content will appear here
02 / Overview

Kling O3 specifications

Facts reviewed:

Developer
Kuaishou / Kling AI
Official name
Kling 3.0 Omni
Official model ID
kling-v3-omni
Name used here and upstream
Kling O3
Official endpoints
POST /text-to-video/kling-3.0 and POST /image-to-video/kling-3.0
Official positioning
All-in-one multimodal input, voice-driven characters, and native audio output with storyboards
Official resolution
720P, 1080P, and 4K
Official clip length
3–15 seconds
Official reference inputs
Up to 7 images or elements without a video; up to 4 with one
Official API price
From 0.6 Units (USD 0.084) per second, by resolution and input combination
Official online date
25 February 2026, with 4K added on 23 April 2026
Platform route
kling-o3

What the official guide documents next to what this page runs

Two documents describe this tier and they overlap without matching: the capability map ticks rows for all six models, and the Omni user guide describes the reference workflow in prose. This page runs one platform route against both.

OFFICIAL

Documented for the official model

  • Reference-driven workflow: text-to-video, image-to-video, first-and-last-frame, multi-image reference, element reference, and video element reference
  • The only model in the capability map whose video reference and multi-image input rows are both marked supported
  • Up to 7 images or elements when no video is supplied, and up to 4 when one is
  • One reference video, 3–15 seconds after the June 2026 expansion, at up to 200 MB and 2K resolution
  • Up to 3 elements in a request, each composable from up to 4 multi-view images
  • Element voice control, which binds a voice to a character so the same subject sounds the same across shots
  • Native audio, a multi-shot switch, and an intelligence shot type for automatic shot splitting
  • 720P, 1080P, and 4K output, with 4K billed at 3.0 Units per second
LIVE PLATFORM

Exposed by this workspace today

  • Text-to-video, image-to-video, and reference-to-video, all limited to one uploaded image
  • A sound switch that defaults to off, matching the official audio parameter rather than the pricing rows
  • A 2–15 second selector, against an official 3–15 second range, defaulting to 2 seconds
  • A resolution selector listing 720p, 1080p, and 4k, with 720p as the default
  • Three aspect ratios: 16:9, 9:16, and 1:1
  • No first-and-last-frame control and no element library, so those official capabilities stay model-side
  • Points quoted before submission on the live estimate

The gap that matters most here is the reference count. The official guide allows up to 7 images or elements without a video and up to 4 with one, and the upstream gateway that carries this route advertises up to 2 input images for its image-to-video entry, while the control on this page accepts one. Three different numbers describe the same capability, so this page reports the official ceiling as a model capability and the single upload as what the route actually takes. The duration default is the second one to read carefully: the official range starts at 3 seconds, and this workspace opens at 2.

Model brief

Kling O3 is Kling 3.0 Omni: two names, one model ID

Kling O3 is the name this platform and its upstream gateway use for the tier Kuaishou registers as Kling 3.0 Omni. The official model ID is kling-v3-omni, the same identifier that appears on the official capability map and on both API endpoints, POST /text-to-video/kling-3.0 and POST /image-to-video/kling-3.0. Searching for Kling O3 and searching for Kling 3.0 Omni land on the same model, and this page states the mapping rather than leaving a reader to guess it.

The short name is not an invention of third-party sites. Kuaishou uses it too: the official comparison article is published at a path containing kling-vs-o3, and its opening argument is that the choice is not about shorthand names but about matching the model to a creative workflow. What the official documentation does not do is use O3 as a model ID. Enumerating every kling- identifier in the official update log returns nine IDs, and o3 is not among them. The tier therefore has an official name, an official ID, and a widely used short form, and this page carries all three.

What makes the tier worth its position is reference handling. The official capability map runs the same rows down all six models, and this is the only model whose video reference row and multi-image input row are both marked supported. The other five, including Kling 3.0 and Kling 3.0 Turbo, are marked unsupported on both rows. Multi-image input is what lets a character, a product, and a location arrive as separate references instead of being compressed into one frame.

The reference surface has limits that are stated numerically and are easy to overstate. The official guide allows up to 7 images or elements in one request when no video is supplied, and up to 4 when a video is. A single reference video is allowed, with a duration that was expanded from 3–10 seconds to 3–15 seconds in June 2026, a file ceiling of 200 MB, and a resolution ceiling of 2K. Elements are limited to 3 per request, and each element can be composed from up to 4 multi-view images. Those four numbers are the whole ceiling, and none of them is a round ten.

Pricing on this tier is structured differently from the rest of the family. Every other Kling model is priced by resolution and, where relevant, by whether native audio is on. This model is priced by input combination as well: with no video input and no native audio it costs 0.6 Units per second at 720P, with native audio 0.8, and with a video input 0.9. The official guide says plainly that whether a video is provided affects generation cost, and the price table shows exactly how.

This page runs the platform route kling-o3 rather than a direct account on Kuaishou’s API. Three deviations follow and are stated separately throughout. The workspace exposes references as a single uploaded image rather than the official ceiling, starts its duration selector at 2 seconds against an official 3, and defaults to that 2-second value rather than the official 5-second default. Resolution is where this route matches the official table exactly, including 4K.

03 / Capabilities

Why the reference-driven tier exists

This is the tier to choose when a subject has to survive several shots, or when the material for a scene arrives as separate pieces rather than as one frame.

01

Video reference, which no other Kling model has

The official capability map marks video reference unsupported for the other five models and supported here, which makes a supplied clip usable as a reference rather than only as an input to edit.

02

Multi-image input

Several reference images can be supplied in one request, so a character, a product, and a setting can each carry their own reference instead of being merged into a single starting frame.

03

Up to 7 references without a video, 4 with one

The official guide states both ceilings explicitly in a single note, and the lower one applies whenever a reference video is part of the request.

04

Element voice control

A voice can be bound to a character element, which the official guide describes as making a subject not only look the same but sound the same across different videos, scenes, and shots.

05

Native audio with a real switch

The official audio parameter takes native or off and defaults to off, and this workspace exposes exactly that control rather than pricing audio in silently.

06

A multi-shot switch and an intelligence shot type

The API exposes multi-shot as a boolean and a shot type for automatic splitting, so a sequence can be divided for you instead of being written shot by shot.

07

4K output with its price published

Resolution reaches 4K at 3.0 Units per second, and unlike the value tier this is a row the official price table actually fills in for this model.

08

Three modes in one workspace

Text-to-video, image-to-video, and reference-to-video all run on the same route here, so a project can move between prompting and referencing without changing pages.

04 / Scene ideas

Work that needs the same subject across several shots

This is the tier to choose when a subject has to survive several shots, or when the material for a scene arrives as separate pieces rather than as one frame.

Characters that must stay recognisable

Bind a character to a reference so faces, wardrobe, and hair survive a cut, a scene change, and a change of camera distance.

Products shot from more than one angle

Supply separate references for a product and a setting, then generate a sequence where the object keeps its geometry through every shot.

Reusing an existing clip as a reference

Use a fragment that already reads correctly as the reference for a new sequence, which is the capability only this tier documents among the six Kling models.

Dialogue scenes where the voice has to match

Attach a voice to the character element so a line spoken in one shot is recognisably the same speaker in the next one.

Storyboards split automatically

Turn the multi-shot switch on and let the intelligence shot type divide the sequence, then adjust the shots that need timing changes.

Delivery work that has to reach 4K

Finish on this tier when the specification names 4K, because the price table prices that resolution here and the value tier cannot reach it at all.

05 / Prompt lab

Reference-driven prompts for this tier

These assume a reference has been supplied. The official multi-shot format still applies: each shot is an index, a duration, and a description separated by semicolons, with one camera movement per shot.

SCENE 01reference to video

Character held across three shots

Attach a character reference and check that identity survives a lighting change and a change of shot size.

View prompt
shot 1, 4, medium shot of the referenced character in a narrow hallway, cool overhead light, holding a folder against the chest, locked-off camera; shot 2, 4, cut to a wider frame as the character turns and walks toward a door, warm light spilling from the room beyond; shot 3, 3, close on the character’s hands opening the folder at a desk. Keep facial features, hair, and wardrobe identical in every shot. Quiet interior ambience, no captions.
SCENE 02reference to video

Product geometry across two angles

Supply the product and the setting as separate references and check that the object keeps its shape.

View prompt
shot 1, 3, the referenced product sits on a matte concrete surface, hard light from the upper left, slow push in; shot 2, 4, the same surface from a forty-five degree angle as a hand rotates the product a quarter turn, reflections sliding across the casing. Hold the casing proportions, the finish, and the surface grain unchanged across both shots. No logos, no captions, restrained commercial look.
SCENE 03reference to video

Existing clip used as a reference

Reuse a fragment that already reads correctly as the reference for a new sequence.

View prompt
Use the supplied clip as the reference for the sequence below. shot 1, 3, a station platform in the same light and colour as the reference, morning haze, fixed camera; shot 2, 4, a train arrives from the right and slows, doors releasing a thin curl of warm air; shot 3, 3, the platform clears as the doors close. Keep the light direction, colour grade, and camera energy of the reference. No added people in the foreground, consistent station ambience.
SCENE 04text to video

Four-shot sequence with sound

Run the multi-shot format with native audio on, then judge whether the shots and the sound agree.

View prompt
shot 1, 3, the referenced character steps out of a lit doorway into rain, neon reflections on wet stone; shot 2, 3, the character walks toward the camera, jacket darkening as it absorbs water; shot 3, 3, a hand pushes a glass door open, interior light spilling out; shot 4, 3, the character sets a wet bag on a counter and looks up. Keep identity and wardrobe stable across all four shots. Rain and street ambience throughout, no music, no captions.
SCENE 05image to video

Single-image move with a sound switch

Use the closest control this workspace has, which can then be lifted later to the 720P resolution with the reference-driven routing if the take holds. Judge identity and geometry at the cheap resolution before opening the switch.

View prompt
Compose a clean starting frame of the uploaded subject against a neutral background with soft frontal light and even exposure, keep the framing itself empty of text, and when you later regenerate the approved take, keep the same framing while the subject offers one short line of on-screen dialogue and the sound switch is turned on. Maintain all other settings between the two runs so the comparison isolates the sound and the prompt change.
COMPARE

Kling 3.0 Omni and Kling 3.0 on the official comparison table

Kuaishou publishes a comparison article for exactly these two models, and the capability map lists both in the same document alongside the other four. Every row below is read from those two official pages.

Kling 3.0 Omni and Kling 3.0 on the official comparison table
ComparisonKling 3.0 Omni (Kling O3)Kling 3.0
Official model IDkling-v3-omnikling-v3
Main workflowReference-driven: text, image, first and last frame, multi-image reference, element reference, video element referencePrompt-led: text, image, first and last frame, first frame plus element reference
Multi-image referenceSupportedNot supported
Video element referenceSupportedNot supported
Element voice controlSupportedNot supported
Element limit per requestUp to 3Not documented
References without a videoUp to 7 images or elementsNot documented
References with a videoUp to 4 images or elementsNot documented
Reference videoOne, 3–15 seconds after the June 2026 expansionNot supported
Official resolution720P, 1080P, 4K720P, 1080P, 4K
API price at 720P, per second0.6 Units without native audio, 0.8 with it, 0.9 with a video input0.6 Units without native audio, 0.9 with it
API price at 4K, per second3.0 Units (USD 0.42)3.0 Units (USD 0.42)
Official online date25 February 202625 February 2026
LIVE PRICING

How Kling O3 is billed, split by input combination

Kuaishou prices this tier by the second of generated video, with the row depending on the input rather than on the model alone. Without a video input the rate is 0.6 Units per second at 720P, which the same table converts to USD 0.084, or 0.8 Units with native audio. With a video input it is 0.9 Units at 720P, or USD 0.126. At 1080P the matching figures are 0.8, 1.0, and 1.2 Units, and at 4K the rate is 3.0 Units per second. The official guide states plainly that whether or not a video is provided affects the generation cost. This site charges platform points instead, so the live estimate in the workspace is the price that applies here.

Open the workspace to see the current estimate
01

Platform points

The workspace quotes points before you submit. Points are set by this site and are not the official Units figure.

02

Resolution, duration and references

The official tier documents 720P, 1080P, and 4K from 3 to 15 seconds with up to 7 references. This workspace lists the same three resolutions but exposes 2 to 15 seconds and accepts one uploaded image per run.

06 / Workflow

Setting up a reference-to-video task

  1. 01

    Decide what the reference is carrying

    A character, a product, a location, or a look. Deciding first keeps a run from spending its reference budget on something the prompt could have described in words.

  2. 02

    Upload the reference and describe only what it cannot show

    The reference fixes appearance; the prompt should carry action, camera movement, light change, and sound. Repeating the reference in words tends to fight it.

  3. 03

    Split the sequence into shots and give each one a duration

    Use the official index, duration, and description format with semicolons between shots, and keep the durations summing to the total clip length.

  4. 04

    Turn the sound switch on only when the scene needs it

    The official audio parameter defaults to off and the price table charges more when it is on, so a silent test pass costs less than a scored one.

  5. 05

    Check identity and geometry before raising resolution

    Judge the draft at 720p, then re-run the approved direction at 1080p or 4K, where the per-second rate rises to 3.0 Units at 4K.

Review before publishing

Reference-tier boundaries, naming, and disclosure

A reference reduces drift but does not remove it. Identity can still shift in profile, geometry can wobble on hard surfaces, on-screen text can come back misspelled, and audio can drift from the picture. The reference is a constraint the model is asked to honour, not a guarantee that it will. Every draft is review material, and the full clip with its complete soundtrack should be watched before publication.

The naming layer on this page is a boundary in its own right. Kling O3 is the short form this platform and its upstream gateway use; the official name is Kling 3.0 Omni and the official model ID is kling-v3-omni. The official comparison article argues that the choice between the generation’s models is not about shorthand names, which is consistent with O3 never appearing as an identifier in the official model list. If you are wiring an API call, kling-v3-omni and the two kling-3.0 endpoints are what belong in the request, not the short form.

The reference ceilings are stated in numbers and should be quoted exactly. Up to 7 images or elements when no video is supplied, up to 4 when one is, one reference video of 3 to 15 seconds at up to 200 MB and 2K, up to 3 elements per request, and up to 4 multi-view images per element. Third-party pages that advertise ten or more reference images are describing a ceiling the official guide does not contain, and this page does not repeat them. Note also that the ceiling is an official model figure, while this route accepts one uploaded image per run.

Price and availability statements here are dated. The per-second figures come from the official API price table read on 13 September 2026, the reference-video expansion from the official update log dated 17 June 2026, and the tier’s online date from the same log. That log also carries a retirement notice for older Kling models taking effect on 15 September 2026, which does not cover this tier but sits close to this page’s review date and should be re-checked rather than assumed. Upload only material you are authorised to use, and do not use a real person’s likeness or voice without permission.

Sources

Kling documents behind the Omni tier

Official model identity, capabilities, limits, and pricing are separated from this platform’s controls, points, and availability. Where the two differ, the difference is stated rather than smoothed over.

07 / FAQ

Kling O3 questions, answered

Is Kling O3 an official model name?
It is the official’s own short form, but it is not an official model ID. Kuaishou publishes its comparison article at a path containing kling-vs-o3 and argues that the choice is not about shorthand names, while the official update log enumerates nine kling- identifiers and o3 is not one of them. The official name is Kling 3.0 Omni and the official ID is kling-v3-omni.
So is Kling O3 the same as Kling 3.0 Omni?
Yes. This platform and its upstream gateway both use Kling O3 for the tier Kuaishou registers as Kling 3.0 Omni, and the pricing this site applies reproduces the official kling-v3-omni table. Searching either name reaches the same model, which is why this page carries both.
Which model ID goes into an API call?
kling-v3-omni, against the official endpoints POST /text-to-video/kling-3.0 and POST /image-to-video/kling-3.0. This platform’s route name is kling-o3, which is a route label rather than the provider identifier.
How much does Kling O3 cost?
It depends on the input combination, not only on resolution. Without a video input the rate is 0.6 Units per second at 720P, or USD 0.084, and 0.8 Units with native audio on. With a video input it is 0.9 Units at 720P. At 1080P the matching figures are 0.8, 1.0, and 1.2 Units, and 4K is 3.0 Units per second.
Why does this page say the price depends on whether I supply a video?
Because the official guide says so in one sentence, that whether or not a video is provided will affect the generation cost, and the official price table carries a separate row for it. This is the only model in the family priced that way.
How many reference images can I use?
The official guide allows up to 7 images or elements in a request with no video, and up to 4 when a video is included. This workspace exposes a single upload per run, and the upstream gateway that carries the route advertises up to 2 for its image-to-video entry. Treat 7 and 4 as the model ceilings and one as what this page actually accepts.
Does Kling O3 support Motion Brush or video extension?
No. The official capability map marks motion brush unsupported for all six models in the generation, and video extension unsupported for all six as well, along with last-frame-only. Several third-party pages list both as features of this model, and the official matrix contradicts them.
Is there an O3 Pro tier?
No. Some API resellers label an endpoint Kling O3 Pro, but the official model list for this generation contains kling-v3, kling-v3-omni, and kling-3.0-turbo, and no Pro variant of this tier. This page does not use that label.
When was this tier released?
The official update log records Kling 3.0 and Kling 3.0 Omni as going online on 25 February 2026, with 4K support added on 23 April 2026 and the reference-video duration expanded to 3–15 seconds on 17 June 2026. One third-party page dates the release to 4 February 2026, which the official log does not support.
Does it generate audio, and is there a switch?
Native audio is supported and the official API parameter takes native or off with off as the default. This workspace exposes that switch and also defaults it to off, so a silent test pass is the cheaper option because the price table charges more when audio is on.
Can I bind a voice to a character?
Official element voice control allows a voice to be bound to a character element, which the guide describes as making a subject not only look the same but sound the same across videos, scenes, and shots. It is documented for this tier and not for Kling 3.0. This workspace has no element library, so the capability is model-side rather than a control here.
Are the credit prices on reseller pages the official price?
No, and they are not interchangeable with it. Reseller pages publish their own credit figures, some of which are consumption-side numbers from a different price table and some of which are invented. This page cites only the official API price table in Units and United States dollars, and the point estimate in the workspace is this platform’s own price.

Other routes when you do not need references

01

Kling 3.0 Turbo

Compare the value tier, which matches this model on multi-shot and native audio and drops every image reference.

02

Seedance 2.5

Move to a different family when multimodal reference and longer clips are the deciding factor.

03

Vidu Q3 Pro

Compare a reference-to-video route from another provider on the same workspace.

04

Image to Video

Open the general still-frame video workspace with model selection.

Render your next scene

Start from an image, a video or a prompt

Return to the live workspace, attach your reference, and review the current point estimate before generating.

Open Kling O3 workspace