How to Prompt Kling 3.0

Kling 3.0 is Kuaishou's flagship video generation model, released February 7, 2026. It is the result of merging two earlier Kling models into one unified architecture: Kling Video 2.6, which handled text-to-video and image-to-video with motion control, and Kling O1, which focused on visual quality and character consistency. The combination produced a model that does not just generate a clip but plans the shots inside it.

Tell Kling 3.0 what happens in a scene, and it assigns camera angles, generates synchronized dialogue with lip-sync, and keeps characters consistent across every cut, all in one pass. It generates from 3 to 15 seconds at native 4K, and it is available on Aristotto.

What is Kling 3.0?

Kling 3.0 is a text-to-video and image-to-video model built around three core ideas that distinguish it from most competing models.

The first is Multi-Shot generation. Where most models produce a single locked angle per generation, Kling 3.0 can execute up to 6 distinct camera angles in one pass, planning cuts, pacing, and coverage the way a director would.

The second is native audio. Dialogue, ambient sound, and custom sound effects generate in the same pass as the video. Lip-sync is phoneme-level and supports five languages natively: Chinese, English, Japanese, Korean, and Spanish, with additional dialects and accents within those.

The third is Elements 3.0, Kling's reference-based character and object locking system. Feed it a clean reference image and the model holds that subject's appearance, face, clothing, and props consistent across every shot and camera angle in the generation.

Prompt adherence improved significantly from 2.6, meaning the model now follows detailed scene descriptions more reliably rather than interpreting them loosely.

The five-layer prompt structure

Kling 3.0's own documentation confirms the five-layer order that produces the most predictable results. Write in this sequence:

Scene, then Characters, then Action, then Camera, then Audio and Style.

Each layer builds on the one before it. The scene tells the model where everything is. Characters tell it who is there and what they look like. Action tells it what physically happens. Camera tells it how to frame and move. Audio and style close with how it sounds and the overall visual treatment.

Every layer you name is one less thing the model decides on its own. Leave the camera unspecified and you will usually get static framing. Leave audio unspecified and you will get a soundtrack the model invents.

Camera language is the actual lever

This is what separates creators who get predictable results from those who do not. The camera layer is where Kling 3.0 earns its real advantage over other models, and it only works if you speak to it precisely.

Generic direction like "cinematic camera movement" or "the camera moves toward the subject" gives the model almost nothing to anchor on. Specific named moves execute reliably: dolly push-in, pull-back, tracking shot, slow orbit, whip-pan, crane up, FPV fly-through, crash zoom. Put the camera instruction near the beginning of the prompt so the model prioritizes it over the rest of the description.

One move per beat, not per prompt. Stacking "dolly push while panning left and tilting up" produces warping and motion conflict. A slow, single, sustained movement reads far cleaner than a compound one. Orbit shots specifically should stay under roughly 30 degrees of arc, anything wider tends to produce spatial distortion.

For framing terms, precision compounds the camera direction: extreme close-up, medium close-up, full-body shot, establishing wide shot. Pair framing with composition guidance when it matters: "centered on the speaker," "rule of thirds with the subject in the left third."

Multi-Shot and Custom Multi-Shot

In automatic Multi-Shot mode, describe the overall scene, the narrative beats, the dialogue, and the camera coverage you need, and Kling 3.0 plans the shots. It reads the prompt as a director would read a scene and assigns cuts, angles, and timing.

In Custom Multi-Shot mode, you control each shot individually: shot content, duration, framing, perspective, and camera movement all specified per shot. This is the mode for anything where a specific beat needs to land at a specific moment.

For Custom Multi-Shot, write each shot as a labeled block. Define the setting once at the top, then let each shot block describe only what changes. One action per shot block keeps the pacing readable.

Character consistency with Elements 3.0

A clean reference image is the foundation. Upload it before writing the prompt, keep the character's name, clothing, hairstyle, and props consistent in every written description across shots, and let Elements 3.0 hold the visual identity through the camera moves.

Character consistency is most at risk during fast camera moves, extreme framing changes, and any shot where the subject is partially out of frame. For dialogue scenes with multiple characters, name each speaker explicitly when they deliver a line.

Video Element Reference, available in Kling VIDEO 3.0 Omni, accepts a short video clip as a character reference rather than a still image, which is particularly useful when a character's movement and mannerisms need to match an existing source.

Native audio and dialogue

Dialogue in Kling 3.0 prompts uses clear speaker labels. Keep the line, the speaker's identity, and the delivery note together:

[Character A, controlled serious voice]: "Let us stop pretending."

This format names the character, the voice quality, and the exact line as a single unit. The model holds voice identity across a scene when the reference is this specific.

For ambient sound, describe the acoustic environment the same way you would describe the visual one: "indoor studio with a low air-conditioning hum," "outdoor market with distant crowd noise and a vendor call from the left." The model composes ambient sound from scene description, not from mood adjectives.

Negative prompts

Negative prompts belong in every generation that involves fast motion, multiple subjects, or complex camera work. The most reliable negative list for Kling 3.0: motion blur, face distortion, warping, morphing, inconsistent physics, extra limbs, background shifting, color flickering. Add anything shot-specific: no subtitle overlay, no camera shake, no duplicate subjects.

Five prompts that work

The marble statue, Text-to-video, 10 seconds

A smooth and deliberate dolly-in tracking shot approaching a classical marble statue of a graceful male figure standing in a marble museum hall. The camera starts from a medium-wide distance and slowly moves forward toward the statue with cinematic precision. As the dolly-in progresses, the camera simultaneously performs a subtle pan right and a gentle tilt upward, gradually revealing the statue's intricate details, flowing drapery, serene facial expression, and elegant posture from a lower angle to a more heroic low-angle view. The movement is fluid, professional-grade, steady, and perfectly controlled, showcasing masterful camera work. Highly cinematic, realistic lighting with soft natural daylight, subtle god rays, and gentle atmospheric haze. Photorealistic, masterpiece cinematography.

The golden hour chase, Text-to-video, 10 seconds

Scene: a cobblestone alley in an old city at golden hour, long shadows from the low sun, shallow enough to sprint through. Characters: a runner in a linen shirt and dark trousers, moving fast. Action: he rounds a corner at full sprint, skids slightly on the cobblestones, recovers and accelerates out of frame. Camera: fast tracking shot running alongside at shoulder height, then a crash zoom into the face as he rounds the corner, pulling back to a wide establishing shot as he exits. Audio and style: boots on cobblestone, the scrape of the skid, breathing accelerating, no dialogue, warm golden color grade with deep blue shadow.

Negative: motion blur, warping, background shifting.

The dance in the rain, Text-to-video, 12 seconds

Scene: an empty city street at night after rain, neon reflections in the puddles, a light drizzle still falling. Characters: a woman in a red dress, defined consistently, standing in the center of the empty street. Action: she begins to dance slowly, alone, arms out, turning once, letting rain catch in her hair and on the dress fabric. Camera: Shot 1, Wide shot, the full street, her small in the frame. Shot 2, Medium shot, waist-up as she turns. Shot 3, Slow orbit around her at chest height, 20 degrees of arc. Audio and style: rain on asphalt, a distant piano heard from an open window two floors up, no dialogue, warm neon-lit color grade with desaturated everything except the red.

Negative: warping, fabric clipping, unnatural movement, inconsistent character appearance.

The heist entry, Text-to-video, 15 seconds

Scene: a glass office building at night, security lights on, empty corridors visible through the windows. Characters: one figure in black, defined consistently.

Shot 1, 0-4s, Wide establishing shot: The building exterior, the figure crouched on the rooftop edge, city lights behind them.

Shot 2, 4-8s, Medium shot: The figure rappels down the glass face of the building, rope taut, boots finding the window frame.

Shot 3, 8-12s, Close-up: Gloved hands working a glass cutter against the window, the circular score line appearing slowly.

Shot 4, 12-15s, Wide interior shot: The window swings inward, the figure drops silently into the darkened office, pauses, looks left.

Audio and style: rope tension and creak, the high-pitched scratch of the glass cutter, soft boot landing on carpet, tense ambient silence underneath, no dialogue, cold blue exterior light giving way to warm amber interior light.

Negative: face distortion, warping, inconsistent character appearance between shots, camera shake on the interior shot.

The morning ritual, Text-to-video, 12 seconds

Scene: a quiet apartment kitchen at dawn, soft pre-light through the window, everything still. Characters: a woman in a white robe, defined consistently.

Shot 1, 0-3s, Wide shot: The kitchen from the doorway, the woman entering frame from the left, moving to the counter.

Shot 2, 3-6s, Medium shot: Her hands filling a moka pot with ground coffee, tamping it gently, screwing the top on.

Shot 3, 6-9s, Extreme close-up: The moka pot on the lit hob, the coffee beginning to rise through the spout, steam curling upward.

Shot 4, 9-12s, Medium close-up: Her face as she wraps both hands around a ceramic cup, eyes closed, the first breath of morning.

Audio and style: the quiet of the apartment, the soft metallic click of the moka pot, the low hiss of the hob, the first gargle of coffee rising, room tone throughout, no music, warm dawn light color grade.

Negative: warping, inconsistent character appearance between shots, motion blur on the close-up.

Where Kling 3.0 is being used

Kuaishou has reported over 60 million creators, more than 600 million videos generated, and partnerships with over 30,000 enterprise clients at the 3.0 launch point. Film pre-visualization, advertising, and storyboarding are the industries where adoption moved fastest, mainly because the Multi-Shot capability makes it practical to generate full scene coverage rather than individual clips.

One limitation worth naming honestly: independent testing found that identifying and holding specific individuals in complex multi-subject scenes is still unreliable. If a shot depends on a specific recognizable face, test it thoroughly before committing a production to it.

What it costs on Aristotto

Kling 3.0 is available on Aristotto. Credit cost varies by resolution and clip length, so for draft iteration stick to lower resolution and only switch to 4K when the take is locked for delivery.