Kling 4.0 Prompt Guide: Timed Beats, @References and Real Examples

All guides · Kling 4.0 in Fuser

Quick answer: the best way to prompt Kling 4.0 in Fuser is as a timed shot list: give each window of the clip its own action, camera move and sound, give every reference image or clip one job, and draft at 720p before you render the keeper at 1080p.

Kling 4.0 runs 3 to 30 seconds in a single pass, twice the 15-second ceiling of Kling 3.0, so a one-sentence prompt no longer fills the clip. Kling's own Kling 4.0 user guide (14 October 2026) structures prompts as timestamped segments and asks you to assign every @ asset a role. In Fuser, one Kling 4.0 node covers text-to-video, start and end frames, reference images and clips, and an Edit mode for existing footage; it picks the route from what you connect. Every example below is a real prompt we ran on 22 and 23 October 2026, shown with its output.

The Kling 4.0 prompt formula

Kling's guide orders a prompt in four parts. Write them in this order:

1. Roles. What each @Image or @Video defines and what must be kept: "@Image1 is the lead: keep her face, hair and raincoat identical in every shot." Skip this part when nothing is connected.

2. Look. Scene, texture, film stock and light, in concrete words: "35mm, pale blue and white palette, hard noon sun through the front glass." Kling's guide lists "premium look" and "vintage feel" as phrasing to avoid.

3. Timed beats. One line per window, "0-5s: …", each with the action, the camera and how the shot ends.

4. Sound. Ambient sound, effects and any dialogue, with the spoken line in quotation marks and the language named. End with "No background music" if you don't want a score.

Kling expands short prompts on its own and follows detailed ones closely, so the length of the prompt decides who directs the clip. Add detail where it changes the picture; Kling's guide warns that longer is not automatically better.

Timed beats: fill 3 to 30 seconds in one pass

Set the length on the node's Duration slider (3 to 30 seconds, default 5), then give each window its own beat. About one beat per five seconds worked for us. This is the prompt behind the 30-second take on the Kling 4.0 model page:

Timed beats (text-to-video, 30 s, 16:9, 1080p): "One unbroken handheld take, direct-flash 35mm snapshot look, fine grain, white tile and bright window light. 0-5s: from behind, a young pastry cook in a white jacket and a long black apron pipes the last pale-yellow buttercream flowers onto a small white cake on a scrubbed wooden bench in a narrow tiled pantry, steel bowls stacked on the shelves. 5-10s: she puts the piping bag down, turns the cake once on its plate to check it from every side, and straightens one flower with a fingertip. 10-15s: she lifts the cake on its plate, turns and carries it through the cramped kitchen past hanging ladles and a steaming stockpot, the camera following at shoulder height. 15-20s: she shoulders open a back door into a quiet, sunlit courtyard and walks across the flagstones. 20-25s: she sets the cake down on a weathered table laid for one and steps back to look at it. 25-30s: she wipes her hands on the apron and walks back inside, leaving the cake alone on the table in the sun. Sound: piping bag squeeze, a pot lid rattling, the door creak, birds outside."

It came back as one take with no cut our scene detection could find, and each beat landed in its window: piping until about 5 s, the cake checked by 10 s, through the kitchen by 15 s, out of the door at about 17 s, on the table between 20 and 25 s, and the cook heading back inside to finish. One thing slipped: Kling kept the white tile and window light but dropped the "direct-flash 35mm snapshot look" for soft daylight, in this take and in both 10-second drafts before it.

For contrast, we gave Kling the same scene as one sentence at 30 s and 720p:

One line (text-to-video, 30 s, 16:9, 720p): "A young pastry cook in a white jacket pipes pale-yellow buttercream flowers onto a small white cake, carries it through a tiled kitchen and sets it down on a table in a sunny courtyard."

Kling storyboarded it itself: seven shots, cutting at about 4, 7, 10, 14, 18 and 22 s, with close-ups of the piping and of her face, a cake stand and a vase of yellow flowers nobody asked for, and a still shot of the finished cake for the last four seconds. Kling's guide says multi-shot is on by default and that simple prompts are expanded with storyline, blocking and camera moves. That can be a fine edit, but it is Kling's. Timed beats make it yours.

Camera moves Kling 4.0 follows

• Shot size: wide, medium, close-up, extreme close-up, overhead, profile.

• Movement: push-in, pull-back, pan, tilt, tracking shot, orbit, crane up, handheld, locked-off. Kling's Kling 4.0 page names tracking shots, orbits, push-ins and pull-backs.

• Cuts: if you want them, name the shot and its time: "Shot 2 (2-4s): close-up". Kling's guide also times cuts to sound or action.

• Timing: put the move inside its beat ("10-15s: … the camera following at shoulder height") so it starts and stops where you need it.

Orbits were the hardest move in our runs: one start-frame take barely moved at all. Naming how far the camera travels and what should come into view helped. This is the prompt behind the orbit on the model page:

Orbit (image-to-video from a still of the ring, 5 s, 1080p): "Orbiting camera move: the camera travels in a wide, continuous arc around the giant cloudy translucent inflatable ring lying on the highland moor, moving at least a half circle so we see it from the side and then from the far side, the moor and the distant hills sweeping past behind it with clear parallax. The late sun slides a bright specular highlight across its textured membrane as the angle changes. Russet heather and dry bracken shiver in the wind; the pale sky stays wide. Sound: moorland wind, the membrane creaking softly."

It gave a clear arc with real parallax, though short of the half circle we asked for. When words don't get the whole move, show it: a reference clip can carry the camera, as below.

@Image and @Video: give every reference one job

Connect images to Reference Images or clips to Reference Videos and the node switches to reference-to-video. Cite them in the order they are connected: @Image1, @Image2 … and @Video1, @Video2 …. The node takes up to 10 images, counting a start image, and up to 5 clips of 3 to 30 seconds each, 30 seconds combined. Kling's guide asks for a role per asset, such as "@Image1: preserve the character's appearance" and "@Video1: reference the action pacing", and says results are less predictable without them.

Character lock with scripted cuts (reference-to-video, @Image1 = one portrait, 8 s, 1:1, 1080p): "@Image1 is the lead: the same young woman with dark shoulder-length hair, dark lipstick and a wet translucent raincoat over dark clothes. Keep her face, hair and raincoat identical in every shot. Four hard cuts, slow-shutter motion blur, pale blue and white palette, hard noon sun through the laundromat's front glass. Shot 1 (0-2s): wide, she pushes the glass door open from the street and shakes the rain off her coat. Shot 2 (2-4s): close-up, she leans on a stainless steel washer and watches the drum turn, her face reflected in its round window. Shot 3 (4-6s): medium, she pulls warm white sheets from a dryer and presses her face into them. Shot 4 (6-8s): she spins once in the aisle between the machines, the raincoat flaring and blurring in the sunlight."

The four cuts landed at 2, 4 and 6 s, as scripted, and the face, hair and raincoat held through all four shots: the door, the washer window, the sheets and the spin.

A reference clip can lend its camera instead of its content. We fed back our own five-second orbit of the inflatable ring as @Video1 and a still of a different object as @Image1:

Borrowed camera move (reference-to-video, @Image1 = a still of the pod, @Video1 = our 5 s orbit clip, 5 s, aspect Auto, 720p): "@Video1: copy only its camera move, a wide, continuous orbit at the same height and speed; ignore its subject and setting. @Image1: the subject and the setting, the giant champagne-coloured satin inflatable pod on the gravel plain with its neon coral circle and the tiny yellow star above it; keep its shape, colours and the hard midday sun exactly. The camera orbits the pod as in @Video1, from its side to its back, the gravel plain and the horizon sweeping past behind it. Sound: wind and the soft flutter of vinyl. No background music."

It took the orbit and nothing else. The camera travels round the pod, the coral circle slides to the side and out of view, and the star stays on top; no heather or hills came across from the moor. Two notes. Auto returned 16:9 although both inputs were portrait, so set the aspect ratio yourself when the format matters. And reference clips add their own length to the cost of the run.

Start and end frames

Connect a Start Image and the node runs image-to-video; the output follows the image's shape, so the aspect ratio control disappears. An End Image is optional, needs a start image, and is not used once reference media is connected. With frames attached, describe the motion between them rather than the picture.

Reveal between two frames (image-to-video, start = a tight crop of the end frame, end = the full photo, 5 s, 1080p): "A slow, steady pull-back. It opens close on the hollow torso of a pale plastic flamingo prototype, its microcontrollers and rainbow ribbon cables visible, and the camera glides back until the whole robot flamingo stands in side profile on cobalt blue floor mats in a bright glass-roofed atrium. Its head turns slightly and a servo in the leg twitches. Soft daylight, quiet hum of servos."

Cropping the start frame out of the end frame gave Kling a path it could not miss: it pulled back from the wiring to the full prototype on the mats, in the same light, and the 3:4 photos gave a 3:4 clip.

Edit mode: change the clip, keep the motion

Set Mode to Edit and connect a Video of 3 to 30 seconds, at least 700 px per side, 24 to 60 fps and up to 300 MB. The edit keeps the source's length. Kling's guide gives two rules: start with "Edit @Video1", because "Reference @Video1" may be read as a request for a new generation, and add "Keep everything else unchanged" with a list of what to preserve.

Season swap (Edit mode, source = our 5 s 9:16 clip, 720p): "Edit @Video1: turn the season to deep winter. Fresh snow covers the lawns, roofs, hedges and parked cars, the sky is white and overcast, and his breath shows in the cold air. Keep everything else unchanged: the boy, his face, hair and pale blue and white windbreaker, his jump, the yo-yo, his run and the camera move."

The jump, the yo-yo throw and the run line up with the source frame for frame, and snow now covers the lawns, the wall, the sidewalk, the roofs and the cars. The windbreaker came back a shade bluer. The output ran 5 s, like the source. For more edit patterns, see how to edit a video with text prompts.

Draft at 720p, finish at 1080p

Resolution defaults to 1080p on the node; switch it to 720p while you work out timing and framing. A 720p second costs about two-thirds of a 1080p second, and Fuser shows the cost on the node before you run it.

Plan for this: a 1080p re-run is a new take, not an upscale of the draft. Of the ten clips we re-rendered at 1080p for the model page, two drifted far enough from the approved 720p draft that we ran them a second time. One changed the dancer's face and cut to two seconds of falling water; the other lit the cinema through a side window instead of the projection booth. Lock what must survive with a start frame or reference images, and budget a second 1080p take for clips that depend on a specific look.

Fix what goes wrong

• Cuts you didn't ask for: multi-shot is on by default. One of our 1080p takes of a dancer cut away to water for about two seconds. Open the prompt with "One unbroken take" and keep the action in one place, as in the 30-second cook prompt, which held; Kling's own example adds "No cuts, no jump cuts".

• A music bed: Kling 4.0 always generates sound with the picture, and the node has no audio switch. Kling's guide calls "no background music" one of only two negative instructions that are truly necessary; the other is "no subtitles". Then list the sounds you do want.

• Text on screen: Kling's guide says text stays stable through camera and character movement. In our test it did: the sign, the chalkboard and the subtitle were spelled correctly in every frame. The timing was loose: we asked for the subtitle from 2 to 4 s and it showed from about 0.5 to 4.5 s. Check spelling and timing on every take. When lettering must be exact, lay it over the clip as a text layer in Fuser's Compositor instead.

• Action that doesn't fill a long clip: a one-line prompt on a long clip leaves the pacing to Kling; ours ended on four seconds of a still cake. Write a beat for every five seconds or so.

• Negatives in general: there is no negative prompt field. Kling's guide says negative instructions can increase errors, so describe what you want and keep "no …" for music, subtitles and the one thing that keeps going wrong.

Text test (text-to-video, 5 s, 16:9, 720p): "Slow sideways dolly along a pale mint-green ferry ticket kiosk on a stone quay at hard noon, the sea bright behind it. Painted on the kiosk in large white capitals: "HARBOUR FERRY". A small black chalkboard on the counter reads "NEXT BOAT 4:15" in white chalk. A ticket seller in a navy cardigan slides a paper ticket across the counter. 2-4 s: a subtitle in small white sans-serif, centred at the bottom of the frame: "Last crossing of the summer." Sound: gulls, water slapping the quay. No background music."

Kling 4.0 settings in Fuser

The Kling 4.0 node exposes:

• Mode: Generate or Edit.

• Prompt: up to 2,500 characters; cite inputs as @Image1, @Video1.

• Video: the clip to change, in Edit mode.

• Start Image and End Image: image-to-video; the end image needs a start image.

• Reference Images and Reference Videos: up to 10 images and 5 clips.

• Duration: 3 to 30 seconds.

• Aspect Ratio: 16:9, 9:16, 1:1 or 21:9, plus Auto when references are connected.

• Resolution: 720p or 1080p.

Build it as a workflow

Kling 4.0 does its best work from inputs you control. In Fuser, generate the lead or the first frame with an image model such as GPT Image, wire it into the Kling 4.0 node as @Image1 or the start image, and keep the timed prompt in its own Text node so a 720p draft node and a 1080p final node can share it. When the take is right but the season, outfit or set is wrong, send the clip to a second Kling 4.0 node in Edit mode on the same canvas. For prompt patterns on the earlier models, see the Kling 3.0 prompt guide and the Kling O3 prompt guide. Still choosing a model? See how Kling 4.0 compares with Seedance 2.5.