Multi-Shot Action Scenes with Kling 4.0: Long Takes, Keyframes and Extends

Action is the hardest thing to fake in an AI short drama. A chase or a near miss needs momentum across several shots: the character has to keep the same face, the same jacket and the same direction of travel while the camera changes angle. A good-looking single clip is not enough; the shots have to cut together.

Kling 4.0 is a strong fit for this because its text-, image- and reference-to-video models each generate any length from 3 to 30 seconds in one pass, with synchronized stereo audio, and the image-to-video model accepts in-between keyframes. This guide shows four ways to build multi-shot action scenes with Kling 4.0 and how to combine them.

A note on terms before we start: the Kling 4.0 model pages do not list a separate “multi-shot” switch. In this guide a multi-shot scene means a sequence you build from several directed beats, either inside one long take or across several generations that cut together.

The Kling 4.0 models you will use

Model page	Inputs	What it is for in an action scene
Text-to-Video	Prompt; 16:9, 9:16, 1:1 or 21:9	Long single takes with several beats and camera moves
Image-to-Video	First frame, optional last frame, up to 8 in-between keyframes (10 images in total)	Pinning the poses of a stunt in order
Reference-to-Video	Up to 10 images, 5 videos, 3 audios	Keeping the hero consistent; borrowing a camera move from a reference clip
Video Extend	A video, a prompt, optional target last frame	Continuing the action from the last frame for 3 to 30 more seconds

All four output 720p or 1080p. List prices on the model pages when this post was written: $0.157 per second at 720p and $0.235 per second at 1080p, so a 5-second 720p clip is $0.785. Reference-to-video with reference videos is priced differently; see the model page. For a cost breakdown per usable clip, read Kling 4.0 pricing.

Plan the sequence before generating

Write the scene as a beat list with screen direction. Our example is from the original series Night Bloom: Jun, a courier in a yellow rain jacket, chases a delivery van that drove off with the wrong package.

Beat	Action	Camera	Direction of travel
1	Jun bursts out of the shop door	Wide, static	Left to right
2	He sprints down the wet alley	Handheld tracking from behind	Away from camera
3	He vaults a stack of crates	Low angle, side-on	Left to right
4	He slips, catches a wall, keeps going	Medium, tracking	Left to right
5	The van turns the corner ahead	Over his shoulder	Away from camera

Keeping the direction of travel consistent is what makes separate shots read as one chase. If beat 3 goes right to left, the audience thinks he turned around. Sketch the beats as a row of thumbnails, even stick figures on paper, with an arrow for the direction of travel in each. Direction problems are obvious on paper and expensive to discover after generating.

Method 1: Several beats in one long take

Kling 4.0 Text-to-Video can hold the whole sequence in one generation of up to 30 seconds. Its prompting guide says to name the subject, action, setting and camera, describe the sound, and for longer clips describe how the scene develops from beginning to end. Write the beats in order with the camera move for each:

Night, a narrow wet alley lit by shop signs. A courier in his early 30s with short black hair, a bright yellow rain jacket and a grey delivery bag bursts out of a flower shop door and sprints left to right after a white delivery van. The camera follows him from behind as he runs down the alley, then swings to a low side angle as he vaults a stack of wooden crates. He lands, slips on the wet cobbles, catches himself against a brick wall and keeps running. The camera rises over his shoulder as the van turns the corner far ahead and disappears. Rain, splashing footsteps, his breathing, the van's engine fading.

Settings: aspect_ratio 9:16, resolution 720p, duration 10.

Tips for long takes:

• Give each beat its share of time. Five beats in 10 seconds is about two seconds each. Cramming more actions in usually drops some of them.
• Name the camera change at the moment it happens (“then swings to a low side angle as he vaults”).
• Say the direction of travel once and keep it.
• Describe the sound. Kling 4.0 generates stereo audio with the picture; footsteps and impacts sell the action.

A long take is the cheapest way to get continuous action, because there are no seams to hide. The trade-off is control: if one beat fails, you regenerate all of it.

Method 2: Pin the stunt with keyframes

When a beat has to land exactly (a vault, a landing pose, a final freeze), use Kling 4.0 Image-to-Video with keyframes. You give it a first frame, the poses in between, in order, and optionally a last frame. Together they can be at most 10 images, and the prompt can even be empty.

Make the frames as stills first (from your character reference, so the face and jacket match), then:

{
  "image": "https://example.com/jun-runup.png",
  "keyframes": [
    "https://example.com/jun-hands-on-crates.png",
    "https://example.com/jun-midair.png"
  ],
  "last_image": "https://example.com/jun-landing.png",
  "prompt": "He sprints at the crates, plants both hands and vaults over them, landing low on the wet cobbles. Low side-on camera tracking left to right. Rain, a thud on landing, splashing footsteps.",
  "resolution": "720p",
  "duration": 5,
  "aspect_ratio": "9:16"
}

Allow enough seconds for the motion between keyframes to look physical; a vault squeezed into 3 seconds tends to look like a jump cut. For a fuller walkthrough, see how to use Kling 4.0 keyframes.

Method 3: Chain shots with Video Extend or the last frame

When you want the chase to keep going after a good take, Kling 4.0 Video Extend generates a new 3 to 30 second segment from the last frame of your clip and returns the original and the new segment joined. Prompt the continuation, not the whole scene:

He reaches the end of the alley and skids into a busy night street. The van is stuck at a red light ahead. He weaves between parked scooters toward it. Camera tracks beside him, left to right. Traffic noise, a horn, his breathing.

Settings: resolution 720p (match the source), duration 6. If you know where the extension must end, add a last_image.

To cut to a new angle instead of continuing the same camera, export the last frame of the clip as a still from your editor and use it as the image of a new image-to-video generation with a different camera direction. The frame carries the face, wardrobe and position across the cut.

Method 4: Borrow a camera move with Reference-to-Video

Kling 4.0 Reference-to-Video lets you point at reference images for the hero and at a reference video for the camera. The model page’s prompting guide puts it simply: say what each reference is for, and to reuse a camera move, say so.

The courier from @Image1 runs through the alley from @Image2, following the camera movement of @Video1. He keeps the yellow jacket and grey bag from @Image1. Rain, splashing footsteps, distant traffic.

Settings: aspect_ratio 9:16, resolution 720p, duration 6. Use your own footage or a previous generation as the reference video. Set keep_original_sound only if you want the reference video’s own audio instead of new sound. More on this model in Kling 4.0 reference-to-video.

Putting it together

A practical recipe for a 20-second chase:

1. Generate beats 1 to 2 as one long take (Method 1), 6 seconds.
2. Generate the vault as a keyframed shot (Method 2), 5 seconds.
3. Extend from the landing into the street (Method 3), 6 seconds.
4. Cut to a reaction or a close-up on the van, 3 seconds.

Draft everything at 720p, cut it, and only re-render the approved shots at 1080p if you deliver 1080x1920. Keep the character description identical in every prompt. Our guide to keeping the same character across AI video shots covers that in depth.

Common problems and fixes

• The hero changes mid-sequence. Move that beat to image-to-video with an approved first frame, or use reference-to-video with the character image.
• Actions are skipped. Too many beats for the duration. Add seconds or split the take.
• Direction flips between shots. Restate “left to right” in every prompt and check the first and last frames before cutting.
• Motion looks floaty. Describe weight and contact: “plants both hands”, “lands hard”, “slides on the wet stones”. Add the sound of the impact.
• The action is too small in 9:16. Vertical frames favor tall movement: jumps, falls, stairs, climbing. Keep the subject in the middle third, away from the bottom and right edges where app buttons and captions sit.

For how Kling 4.0 compares with the other two models in this series, see Seedance 2.5 vs Wan 3.0 vs Kling 4.0. The full episode workflow is in how to make an AI short drama.

FAQ

How long can a Kling 4.0 action shot be? Any length from 3 to 30 seconds per generation on text-to-video, image-to-video and reference-to-video. Video Extend adds 3 to 30 seconds per extension.

Does Kling 4.0 have a multi-shot mode? The model pages do not list a separate multi-shot setting. You build multi-shot scenes by directing several beats in one long take, by keyframing image-to-video, or by chaining generations with Video Extend or a shared last frame.

How many keyframes can I use? Image-to-video takes up to 10 images in total across the first frame, the in-between keyframes and the last frame, which leaves up to 8 in-between keyframes when you use both ends.

Does Kling 4.0 generate sound for action scenes? Yes. Kling 4.0 text-to-video, image-to-video, reference-to-video and video extend all list synchronized stereo audio generated with the picture. Describe the footsteps, impacts and ambience you want in the prompt.

What does an action sequence cost? At the list price of $0.157 per second at 720p, a 20-second sequence is $3.14 before retakes, or $4.70 at 1080p.

Ready to direct your first chase? Open the video generator with Kling 4.0, go straight to Kling 4.0 Text-to-Video, or direct individual shots with preset camera moves and film looks in Cinema Studio.