Kling 4 Prompt Guide: References, Keyframes, Audio, and Editing

Kling 4 can work from one short idea or a production-style brief. The right prompt depends less on word count than on the kind of control you need: are you exploring a concept, preserving a character, following a reference movement, fixing visual milestones, generating dialogue, or editing an existing clip?

This guide explains the complete prompting system across Kling 4's main workflows. For specifications and launch status, see the Kling 4 release guide.

Start by Choosing the Right Kling 4 Mode

Kling 4 presents two primary entry points. They solve different problems and cannot be combined within one generation task.

Choose Omni Reference when you want to generate from text, preserve a character or product, borrow motion or camera language, combine several assets, or edit an uploaded video.

Choose Keyframes when your non-negotiable requirement is visual placement in time: start on this image, arrive at that image, or pass through several approved frames in a defined order.

Start with one idea

A good Kling 4 prompt does not need to be long from the first attempt. Begin with the event you want the viewer to see, then add controls only where the model needs less freedom.

A useful starting point is:

An exhausted astronaut finds a tiny green plant growing inside an abandoned lunar base.

Then add details in three passes:

1. Core event: Who or what is present, and what changes?
2. Must-keep decisions: Which identity, composition, movement, or lighting details cannot drift?
3. Production controls: Which references, time ranges, camera moves, audio layers, or edit boundaries matter?

This progression keeps the prompt readable while making the important constraints explicit. It also makes revision easier: if the result fails, you can tell whether the problem came from the idea, the asset roles, or the timeline.

Use a four-block prompt structure

For a controlled Omni Reference task, separate the prompt into four blocks. The labels are not magic commands. They are a way to prevent identity, action, timing, and sound instructions from competing in the same paragraph.

[Asset Roles]
@Image1 defines the character's face, hairstyle, and clothing. Preserve those details.
@Image2 defines the room layout and window light. Preserve the furniture and spatial arrangement.
@Video1 provides the walking rhythm only. Do not copy its people, location, objects, or audio.

[Overall Settings]
16:9, 10 seconds. Natural skin texture, soft afternoon window light, medium lens,
shallow depth of field, realistic motion, stable exposure.

[Sequence of Events]
0-3s: The character enters from the left and walks toward the window.
3-7s: The camera tracks backward at a steady distance as the character slows down.
7-10s: The character stops, turns toward the light, and holds the pose.
Keep the shot continuous and preserve the room layout throughout.

[Audio]
Quiet room tone, soft footsteps, fabric movement, no dialogue, no background music.

1. Asset roles

Give each uploaded asset a specific function. A reference image might control a face, outfit, product geometry, or room. A video might control motion rhythm or camera movement. A voice profile might control the voice of one speaker.

Do not assume that an upload label explains itself. Use the label and the role together, then state what must not transfer. For example, “reference the camera path of @Video1 only” is more precise than “use @Video1.”

2. Overall settings

Replace abstract adjectives with visible or audible properties. Instead of “premium,” describe edge lighting, material response, lens choice, surface texture, and background separation. Instead of “vintage,” specify aged paint, warm incandescent light, worn edges, or film grain.

The official guide makes the same general distinction: concrete descriptions reduce the amount of visual inference the model must perform.

3. Sequence of events

Write actions in the order the viewer should see them. Add time ranges when a longer clip has several beats, and explain how the camera moves between those beats.

A timeline should also say what stays continuous. If the character, room, light direction, or product placement must remain stable, state that rule after the events rather than repeating it in every line.

4. Audio

Separate dialogue, voice characteristics, ambience, music, and sound effects. This prevents a request for a quiet room from conflicting with a request for a music bed or an unclear speaker assignment.

Write text-to-video prompts

Text-to-video works best when the prompt answers four practical questions:

• Subject: Who or what should the viewer follow?
• Scene: Where does the action happen, and what physical details define the location?
• Action: What changes over time, and in what order?
• Treatment: How should the camera, light, texture, sound, and pacing present the action?

Here is a complete example:

A ceramic artist works alone in a small pottery studio during a summer thunderstorm.
The artist centers a hand-built clay bowl on a spinning wheel, wearing a dark linen apron
with rolled sleeves. Rain streaks the tall windows, and cool blue daylight fills the room
while a warm tungsten lamp illuminates the wheel.

12 seconds, 16:9, one continuous shot. Begin with a close view of wet clay under the
artist's hands. The camera slowly pulls backward and slightly to the right, revealing the
studio shelves, tools, and rain-covered windows. The bowl rises gradually as the wheel
turns. Keep the artist's hands, bowl shape, shelf positions, and light direction consistent.

Audio: rain against glass, low thunder in the distance, the soft motor of the wheel, and
subtle clay movement. No dialogue, no subtitles, no background music.

The point is not to use every possible detail. The point is to make the important decisions observable: what is in frame, what moves, how the camera reveals it, and which sound layers belong in the scene.

Use Omni Reference

Omni Reference is the right choice when text alone cannot reliably carry the visual identity, motion, or composition you need. Kling's official materials describe a combined reference budget of up to 15 inputs, including up to 10 images, up to 5 videos with a 30-second total video duration, up to 7 elements, and up to 3 voice profiles. The limits share a total budget, so the category maximums should not be added together.

The same official guide lists image, video, voice-profile, and element limits for its current workflow. Interfaces and integrations can change, so verify the controls shown in your selected workspace before preparing a large batch of assets.

Character and product references

Assign identity references to the details that must survive generation:

@Image1 defines the character's face, hairstyle, and age. Preserve these features.
@Image2 defines the character's coat and accessories. Preserve the fabric, color, and fit.
@Image3 defines the product's shape, label placement, and surface finish. Do not redesign it.

If an asset contains extra information that should not transfer, say so. For example, a product photo can define the bottle geometry without defining the studio background, reflections, or text that happens to appear behind it.

For workflows that start from an existing still, image-to-video generation can be a useful companion path when the main task is to animate one prepared composition rather than coordinate many references.

Action and camera references

A reference video can guide movement without becoming the scene. State exactly which layer to borrow:

Reference the walking rhythm and lateral camera movement of @Video1 only.
Do not copy its characters, location, vehicles, wardrobe, color grade, or audio.

This matters because a video reference contains more information than the action you want. Narrowing its role makes the intended transfer easier to inspect and revise.

Composition and storyboard references

Use an image when the composition is the decision you want to preserve. Use a storyboard or layout reference to describe the order of shots, then tell Kling to render the scene itself rather than reproduce labels, arrows, panels, or planning marks.

A useful rule is to upload the smallest set of references that controls the decisions you care about. More inputs do not automatically create more control. They can also introduce competing instructions.

Use first and last frames

Keyframes are useful when the opening and ending images are already known. Do not describe only the start and end. Describe the path between them.

Use @Image1 as the first frame and @Image2 as the last frame.
The camera moves slowly toward the upright glass bottle, transitioning from the wide studio
composition to the close-up. Soft reflections travel across the bottle as the camera moves.
Keep the bottle vertical, preserve the label, and maintain the same background lighting.
The transition is continuous and physically plausible.

The first frame establishes the opening composition. The last frame establishes the destination. The middle of the prompt explains how the camera, subject, light, and materials get there.

Plan multiple keyframes

Kling 4's official guide lists support for up to 10 keyframe images. Use a keyframe when a visual state matters at a particular moment, not simply because the interface allows more uploads.

0-4s: @Image1. A wide view of the empty train platform at dawn. The camera is locked.
4-8s: Transition toward @Image2. A commuter enters from the far right as the camera begins
      a slow push forward. Keep the platform architecture unchanged.
8-12s: @Image3. The commuter reaches the yellow line and looks down the tracks.
12-16s: @Image4. Warm sunlight reaches the platform. Preserve the character's coat and
        direction of travel while the camera settles into a medium shot.

Useful keyframe milestones include:

• A character entering or leaving the frame
• A change from wide shot to close-up
• A product changing state or position
• A weather or lighting transition
• A transformation that needs a defined intermediate state

Explain the transition between each pair of keyframes. Without that connective instruction, the model has to infer more of the motion and continuity.

Plan a 30-second multi-shot video

Kling 4 supports 3-30 second generation in one-second increments, according to the official user guide. A 30-second prompt should read like a compact shot plan, not one long paragraph of adjectives.

Include five layers:

1. Spatial plan: Where are the subject, camera, and important objects?
2. Shot plan: How many shots are there, and how long is each one?
3. Action: What is the main event in every shot?
4. Transitions: How does one shot lead into the next?
5. Continuity: Which identity, light, direction, and audio rules carry across the cut?

30 seconds, 16:9, three connected shots. Keep the same red delivery bicycle,
rider's yellow rain jacket, wet street reflections, and overcast lighting throughout.

Spatial plan: the rider travels from the left side of the market street toward the station
entrance. Pedestrians move in the same direction as the bicycle, never through it.

Shot 1, 0-8s: Wide tracking shot from the front-left as the rider turns into the street.
Shot 2, 8-20s: Side tracking shot. The rider passes food stalls while rain collects on the
camera lens. Maintain the rider's speed and direction.
Shot 3, 20-30s: The camera moves ahead and rotates toward the station entrance as the rider
slows, parks the bicycle, and looks up at the station sign.

Audio: rain, tires on wet pavement, distant market voices, and one bicycle bell at 23s.
No subtitles, no background music, no sudden change in weather or wardrobe.

Longer duration creates more room for narrative development, but it also gives continuity errors more time to appear. Keep the spatial rules and recurring identity details visible in the prompt.

Prompt dialogue and stereo sound

For dialogue scenes, define the speaker before the line. Include language or accent when it matters, then describe age, tone, pace, emotion, pauses, and whether other characters remain silent.

Dialogue:
Speaker A, a calm woman in her 30s, speaks in a low, measured voice:
"We have one chance to get this right."
Pause for half a second before the final sentence.
Speaker B remains silent and listens, then nods without speaking.

Ambience: quiet control room, low ventilation, faint keyboard sounds.
SFX: one alert tone at the end of the line.
Music: none.
Stereo: Speaker A is centered. The alert tone begins on the right and moves toward center.

Kling's official launch materials describe two-channel stereo audio and improved lip-sync performance. Treat that as a model capability, not a promise that every interface exposes identical audio controls.

For music, ambience, and effects, use separate labels. For spatial sound, describe the movement through the stereo field as an event, such as a vehicle moving from rear-left to front-right.

Write video-editing prompts

When you are changing an uploaded clip, start by identifying the video as the material to edit. A practical structure is:

Edit @Video1.
Change: [the exact element or property to change]
Timing: [when the change happens and how it progresses]
Preserve: [identity, camera, geometry, lighting, audio, duration, and everything else]

Remove an object

Edit @Video1.
Change: Remove the advertising board in the foreground from the first frame onward.
Timing: Keep the removal consistent as the camera pans right.
Preserve: The people, building geometry, shadows, reflections, camera movement, audio,
and original duration. Do not add a replacement object.

Replace a subject or background

Edit @Video1.
Change: Replace the backpack with the backpack shown in @Image1.
Timing: Match its position, scale, movement, contact points, and occlusion throughout.
Preserve: The person's face, clothing, hands, walking pace, background, lighting, shadows,
camera path, audio, and duration.

Change one part of the style

Edit @Video1.
Change: Convert the vehicle into a hand-built clay model with visible fingerprints and
soft matte surfaces.
Timing: Apply the material change gradually during the first two seconds.
Preserve: The vehicle's size, route, wheel motion, camera movement, background, lighting,
shadows, audio, and all other objects in the scene.

The Preserve line is not filler. It limits unintended edits and gives you a concrete checklist when you review the result.

If your project requires a final quality pass after generation, video upscaling is a separate finishing step. It should not be used to hide a prompt problem that needs a new generation.

Use negative instructions sparingly

Describe the desired image or action directly whenever possible. “Show the character in profile while speaking on the phone” is usually clearer than “Do not show the character's face.”

Keep negative instructions for specific failure boundaries, such as:

• No subtitles
• No background music
• No extra characters
• Do not copy the people, location, or audio from the motion reference
• Preserve every other object and the original duration during an edit

Long lists of “no” statements can compete with the positive scene description. Use them to close a known failure mode, not to describe the entire video by exclusion.

Revise in the right order

When a result misses the brief, change one control layer at a time:

1. Check the labels. Make sure each @Image, @Video, element, or voice profile refers to the intended upload.
2. Clarify asset roles. State whether a reference controls identity, composition, motion, camera, or sound.
3. Replace abstract adjectives. Convert “cinematic” or “premium” into light, material, lens, camera, and texture details.
4. Repair the timeline. Add time ranges and define how the action reaches the next state.
5. Strengthen continuity. Name the identity, direction, light, geometry, and sound that must remain stable.
6. Reduce edit scope. Change one subject, object, or style layer before requesting several unrelated edits.
7. Remove conflicts. Delete repeated or contradictory instructions rather than adding more text.

This order makes the next generation easier to diagnose. It also avoids turning every failed clip into a longer, less coherent prompt.

Check AI disclosure before publishing

If a Kling 4 output looks realistic and depicts a person, place, event, or scene that did not happen as shown, review the platform's disclosure rules before publishing. YouTube's current guidance says creators must disclose realistic AI-generated or meaningfully altered content, including realistic scenes that never occurred and footage that makes a real person appear to say or do something they did not do.

That disclosure step is separate from prompt quality. A well-written prompt can improve the result, but it does not change the responsibility to label synthetic or altered media when the destination platform requires it.

FAQ

What is the best structure for a Kling 4 prompt?

For controlled generation, use four blocks: asset roles, overall settings, sequence of events, and audio. For a simple idea, start with one sentence and add only the constraints that affect the result.

Should I use Omni Reference or Keyframes?

Use Omni Reference when you need to combine images, videos, elements, voice profiles, or a creative concept. Use Keyframes when specific images must appear at specific points in the timeline. The two creation entry points have different controls and cannot be treated as interchangeable.

How do I reference an uploaded file?

Use the label assigned by the workspace, such as @Image1 or @Video1, and state what that asset controls. “@Video1 controls walking rhythm only” is more useful than naming the file without a role.

How do I write a Kling 4 video-editing prompt?

Start with Edit @Video1. Then separate the request into Change, Timing, and Preserve. Specify the exact element to change and list the identity, camera, lighting, audio, and duration that must remain stable.

Can Kling 4 generate dialogue, music, and sound effects?

Kling's official materials describe generated audio, two-channel stereo, dialogue, and improved lip sync. Describe dialogue, voice, ambience, music, and effects as separate layers, and check the controls available in your selected interface.

Are longer prompts always better?

No. The official guide lists prompt input up to 8,000 tokens, but that is a capacity limit, not a target. A shorter prompt with clear roles, timing, and preservation rules is often easier to follow than a long prompt full of repeated adjectives.