Kling 4.0 is Kling AI's next video model family. The main model is specified for up to 30 seconds of video and 10 keyframes; Flash tops out at 20 seconds and lacks keyframe controls. The practical appeal is greater control over references, timing and performance across a sequence.
This guide separates announced capabilities from account availability, then shows how to prepare references, write prompts and check whether a result is usable. The specifications and their sources are compared below.
You can explore examples and start a draft in ReelMuse's Kling video generator. Choose an available model, begin with a prompt or image, and use the shot-planning method below to refine the result.
Reviewed October 8, 2026. This is a documentation-based guide, not a hands-on benchmark. The illustrations come from Kling's official examples; the original prompts below are proposed starting points.
What is available, and what is still coming?
Kling's September 30 announcement describes the main model entering early access, with a broader October rollout. Flash access began September 28 for Ultra Yearly members. At this review date, the feature page still labels the full model coming soon and Flash a limited rollout. Check availability in your own account; these announcements do not establish access through every reseller or API. See the official launch announcement and current feature page.
| Capability | Main model: published specifications | Flash: published specifications |
|---|---|---|
| Generation length | 3–30 seconds | Up to 20 seconds |
| Output resolution | 720p, 1080p, 4K | 720p |
| Generation modes | Text-to-video, image-to-video, Omni Reference | Text-to-video, image-to-video, Omni Reference |
| First and last frames | Supported | Not supported |
| Multiple keyframes | Up to 10 | Not supported |
| Dynamic range | 10-bit HDR at 1080p/4K announced as coming soon | 8-bit SDR |
| Longer extensions | Up to two minutes announced as coming soon | Do not assume the same capability |
The main-model specifications are documented in Kling's model release notes; the Flash restrictions appear in the official FAQ. For HDR and two-minute extensions, this guide follows the explicit coming-soon status in the launch announcement, even where comparison tables omit that qualification.
Treat Flash as a separate configuration. It can help explore a scene or reference combination, but a Flash result cannot validate a multi-keyframe sequence when that control is unavailable there. Before buying access, check the exact model name, required input controls, export settings and charge shown for your task.
What changes from Kling 3?
The previous generation already offered native audio, multi-shot workflows and 4K output. The larger changes are the longer generation window, intermediate keyframe control, a higher reference allowance and broader editing controls. Kling also describes improvements to dynamic motion and audiovisual coherence. Those are vendor claims, not measured success rates. The official version comparison provides the baseline.
These changes matter differently depending on the job:
- A product demonstration needs a stable object through contact, movement and a final presentation frame.
- A recurring character series needs recognizable identity across angles, locations and separate generations.
- A dialogue scene needs consistent speaker assignment as well as believable faces and sound.
- A planned sequence needs the important events to happen in the right order, with useful time between them.
A longer duration helps only if it serves the sequence. A failed handoff halfway through a 30-second clip can make the whole take unusable. If a project will be edited into short shots anyway, generate around those shot boundaries instead of forcing every idea into one take.
Give every reference a specific job
The published reference budget is 15 assets in total, with subtype limits of 10 images, five videos and seven subjects. Video references have a combined 30-second limit, including video-based subjects. These are overlapping limits, not an allowance of 22 separate uploads. Kling also documents prompts up to 8,000 tokens; tokens are not characters. Official reference and prompt specifications.
More inputs create more opportunities for disagreement. Start with the smallest set that explains the shot, and name the responsibility of each input. This reference-role approach is also central to PhotoGPT's prompting guide.
| Input | The question it should answer | A useful instruction |
|---|---|---|
| Character image or saved subject | Who is in the scene? | Preserve this person's face and jacket. |
| Product image | What must retain its shape and materials? | Use this radio's proportions and control layout. |
| Environment image | Where does the scene take place? | Use this room's layout and window position. |
| Motion reference video | How should the action unfold? | Follow the hand movement, without copying the performer's appearance. |
| Camera or previs video | How should the camera and subjects move through space? | Preserve the low tracking angle and walking path. |
| Storyboard or keyframe inputs | Which visual moments must occur? | Follow the intended sequence and describe the transitions. |
Keep appearance and movement instructions separate. Otherwise, a clip supplied for camera motion may also influence the character or setting. If two images disagree about clothing, specify which one controls clothing—or remove the conflicting image.
A rough 3D video can carry precise staging
Previsualization, often shortened to previs, is a rough version of a shot made before the final imagery. A simple 3D animation can communicate movement, object placement and camera position without realistic textures.
Official example: the first image establishes staging; the second shows the result with a different appearance. These stills illustrate the relationship between inputs and output, not motion accuracy across the complete clip.
For your own version, explain what should survive the transformation: the walking path, the hand holding the case, the camera height and the timing. Then assign character and environment appearance to separate references. Judge the result against those individual responsibilities, rather than an overall impression of similarity.
Keep character identity separate from scene direction
A reusable character reference should resolve ambiguity before generation starts. Use a clear front view, a three-quarter or profile view when the shot needs one, and a view that shows the outfit. Keep hairstyle, distinguishing features and costume consistent across the set.
An official character sheet separates identity and costume details from the later scene. Multiple views here describe one character, not a cast of several.
The useful distinction is between identity, wardrobe and role. A face can stay recognizable while the jacket changes. Two characters can each look correct while their actions or dialogue are swapped. Check all three separately.
Before a full sequence, try short tests that stress different weaknesses: a head turn, a brief occlusion followed by reappearance, and the same character in a new setting. Reuse the approved reference set for each shot. If identity fails only after a turn, fix that angle or simplify the motion before expanding the scene.
Atlas Cloud's character-consistency article offers a useful reference-pack and testing method. Its demonstrated workflow uses the earlier Kling O3 model, despite the article's title. Treat it as a preparation method, not evidence of this generation's measured performance.
Storyboards, first/last frames and keyframes solve different problems
These inputs all involve images, but they express different intentions.
A storyboard describes progression and composition. A first/last-frame pair constrains the start and finish. Multiple keyframes add intermediate visual targets. Choose the control that matches the uncertainty in your shot: overall structure, endpoints or specific moments in between.
The Chinese heading specifies eight beats in one continuous shot: the camera follows the character from the moon's surface into a tunnel and back outside. Eight panels describe the progression without requiring eight cuts.
If a storyboard should become one continuous take, say so and explain how the camera moves between compositions. If it should become several shots, identify the cuts. A wide view followed by a close-up could mean a dolly move, a zoom or an edit; leaving that choice open invites an unintended result.
This official overview collects five visual checkpoints in one image. An overview collage is not the same input as five individually assigned timeline keyframes.
Plan the transitions as carefully as the target frames. If the opening shows a closed case and the ending shows it open, specify who opens it and when. Avoid requesting an immediate relocation or a new object without an action or cut that explains the change. This is the practical implication of PhotoGPT's distinction between reference layouts and temporal instructions.
Use fewer, compatible milestones first. Once that sequence works, add a checkpoint where an important beat is drifting. Filling every available slot gives you more constraints to reconcile.
How to write a prompt that can be checked
A useful prompt tells you what to inspect afterward. Separate the shot into subject, action, camera, timing, sound and fixed details. Put the most consequential requirements first; use duration and aspect-ratio controls in the interface as well as prose where appropriate.
Subject and references: who or what appears; what each asset controls.
Action: one primary action, with a clear start and finish.
Camera: framing and one defined movement, or a locked position.
Timing: the order of important beats and any intended cuts.
Sound: exact dialogue, speaker, delivery and necessary ambience.
Fixed details: the few properties that must survive the shot.
The examples below are original briefs, not tested outputs. Use an Omni Reference workflow that supports the listed inputs, attach the assets and set the duration in the interface. Replace the descriptive asset names with your actual reference labels; they are not universal API syntax. Timing instructions express your intention, not a guarantee of frame-accurate execution.
To practice this structure in ReelMuse's video tool, choose text-to-video for a written brief or image-to-video for an approved starting image. Paste the relevant prompt, adapt its references to the selected model's inputs, and choose the supported duration and aspect ratio. Keep the first attempt focused on one action, then use the checks below to decide what to change.
Product shot: make the action physically specific
Inputs: one product image and one room image. Duration: eight seconds.
Use the product reference for the teal tabletop radio's shape,
speaker grille and control layout. Use the room reference only
for the wooden shelf and morning window light.
Eight seconds, one continuous close shot. The radio rests on
the shelf. A right hand enters from screen right, grips the
volume knob, rotates it clockwise once, releases it and leaves.
The radio stays in place. The camera remains locked.
A quiet mechanical click, followed by soft radio static.
Keep the casing color and knob positions consistent.
This gives you explicit failure conditions: the wrong hand, a moving radio, a knob that changes position, or a hand passing through the object. Fix the failing requirement before adding dramatic camera motion. If product typography must be exact, allow for a separate finishing step and inspect every visible label.
Character shot: expose the identity risk early
Inputs: one character sheet and one workshop image. Duration: ten seconds.
The character sheet shows one bicycle mechanic. Preserve the
same face, short dark hair and mustard work jacket in every view.
Use the workshop image for the location only.
Ten seconds, medium shot at eye level. The mechanic examines
a bicycle wheel, turns into profile to place it on a stand,
then looks back toward the camera with a small, tired smile.
The camera stays still. Keep the wheel in the same hand until
it rests on the stand. Quiet workshop ambience; no dialogue.
The profile turn tests recognition; the wheel tests object ownership and contact. If both are unreliable together, first test the turn without the prop. Change one source of difficulty at a time so the next result teaches you something.
Dialogue shot: assign the voice and the line
Inputs: separate reference images for the clerk and traveler. Duration: twelve seconds.
Use reference A for the station clerk and reference B for the
traveler. Keep their identities and outfits separate.
Twelve seconds, a locked medium two-shot at a quiet ticket desk.
The clerk stands on the left; the traveler stands on the right.
The traveler places a folded map on the desk.
Traveler, in English, quietly: "Does the last train stop here?"
A short pause. Clerk, in English, matter-of-fact:
"Tonight, it does."
Only the assigned speaker talks during each line. Low station
room tone; no music. The traveler listens during the reply.
Specify a voice reference separately when the chosen interface supports it. A portrait defines appearance; it does not uniquely define a voice. Leave enough time for the line, pause and reaction rather than packing a long script into a short clip.
Judge motion, dialogue and audio separately
Kling advertises stereo sound and improved lip sync. Evaluate those as separate properties: two audio channels do not establish that words match mouth movements, and a believable voice does not establish that it belongs to the correct character. Official audiovisual comparison.
A dialogue still can show framing and expression. Speaker timing, lip sync and spatial sound require watching and listening to the full result.
Use three review passes: silent playback for visual continuity, audio alone for words and delivery, then synchronized playback for timing. The wider production issue—whether movement, character and sound remain coherent—is discussed in IT Supply Chain's overview.
| Check | Inspect this moment | Common failure to reject |
|---|---|---|
| Identity | Profile turn or reappearance after occlusion | Face, hair or wardrobe changes |
| Contact | A foot landing or a hand releasing a prop | Sliding feet, penetration or floating objects |
| Object continuity | Before and after a handoff or cut | Shape, number or ownership changes |
| Camera continuity | A tracking move past a fixed landmark | Background geometry bends or position jumps |
| Dialogue | Start and end of each spoken line | Wrong speaker, missing words or mouth drift |
| Audio | Pauses, movement and the final frame | Unwanted music, discontinuities or clipped endings |
These checks are a proposed acceptance method, not a claim about how frequently the model fails. Define which errors are fatal for your project before generating. A tiny costume variation may pass in an atmospheric montage and fail in a product campaign.
4K, 10-bit and HDR describe different things
Resolution describes image dimensions. Bit depth describes numerical precision: 8-bit has 256 possible code values per component; 10-bit has 1,024. HDR also involves how brightness and color are represented and displayed. A 10-bit file is not automatically HDR, and a high-resolution preview does not prove native detail. ITU-R BT.2100 defines HDR television image parameters.
For an actual delivery, inspect the downloaded original and the destination's requirements. A social platform may process it differently from your editing application. The compressed illustrations in this article cannot demonstrate an HDR export.
If you have FFmpeg installed, ffprobe can report video stream metadata:
ffprobe -v error -select_streams v:0 \
-show_entries stream=codec_name,width,height,pix_fmt,color_primaries,color_transfer,color_space \
-of json output.mp4
Use this to check dimensions, pixel format and color signaling. Metadata is a first check, not a complete validation of the grade or how it will appear on an HDR display. Resolve the delivery format before paying for a larger batch.
Budget for usable footage, including retries
A generation price tells you the cost of an attempt. Production cost depends on how much of that attempt survives review. Keep a simple record of the model, settings, charge, result and seconds actually retained.
Generation cost per usable second =
total generation credits spent / seconds retained in the edit
For illustration, six attempts at 20 credits each cost 120 credits. If you retain 12 seconds across the results, the generation cost is 10 credits per usable second. These are hypothetical numbers, not Kling's pricing or a measured acceptance rate. Editing labor and other production costs sit outside this calculation.
A practical workflow is to approve the reference pack, test the difficult action briefly, then expand timing and output quality. Use Flash when its supported controls answer the question you are testing. Switch to the intended delivery model before validating a feature that Flash lacks.
When comparing models, hold the brief, assets, duration and acceptance criteria constant where possible. Record different settings when you cannot match them. Count rejected attempts as well as successful ones; one attractive result cannot establish which model is more economical. TopMediai's review also frames retry cost as a practical selection issue.
Common questions before choosing a workflow
Is Kling 4.0 the same as Kling 4K?
No. The version number identifies the model family; 4K describes resolution. The older Kling 3 series also supports 4K, so verify the selected model name separately from the export size. Official comparison.
Can I generate a two-minute video in one pass?
Do not plan around that. The published main-model generation limit is 30 seconds; two-minute length refers to an announced extension workflow that is still labeled coming soon in the cited release information. Official feature FAQ.
Is it free, and is there a public API?
The cited rollout announcement does not establish universal free access. Check your account's model selector and charge before generating. For API access, confirm the exact model identifier and supported parameters in the provider's documentation. Access to Kling's consumer application does not, by itself, establish API availability. Start with the official API documentation.
How do I try this workflow on ReelMuse?
Open the ReelMuse Kling video generator, choose a model and start with one short shot. Use a source image when appearance is already decided, or a text prompt when you are exploring a scene. Select the available output settings, review the credit cost and generate a first version. Compare it against the acceptance checklist before adding more movement or a longer sequence.
At this review date, the tool defaults to Kling 3.0 Omni; native Kling 4 is not yet connected. Duration, reference limits, resolution and credit cost follow the model selected in the generator. The announced specifications in this guide do not unlock those settings in the current tool.
Sources and example provenance
Official documentation takes priority for specifications and release status. The supplied analyses add workflow context: ImagineToVideo distinguishes the version announcement from older 4K features; Crun separates announced controls from pending features; ImagineArt presents an announcement-based comparison rather than a completed hands-on benchmark. The prompting, consistency and production articles are linked beside the relevant discussion above.
Illustrations are existing ReelMuse CDN copies of examples from Kling's model release showcase. They illustrate reference design and scene structure, and were not generated from this article's original prompts.
