Blog · In partnership with Higgsfield

Kling 4.0: Full Breakdown
Cinematic Video

I ran the same tests through Kling 4.0, Wan 3.0 and Seedance 2.5 on Higgsfield. This is the full Kling 4.0 workflow I use to build cinematic shots from one recorded performance, plus where it wins, where it doesn't, and what it actually costs.

Record→Cast→Set up→Prompt→Check→Scale→Finish

Kling 4.0Made with Kling 4.0 on Higgsfield.

Best for

Cinematic shots

Key trick

Swap the cast, keep the performance

Cost (1080p)

7.5 credits / sec

Part 1 · The verdict, up front

Three models, same prompts, same settings, four kinds of scenes: cinematic, fast action, motion design and character work. Here's where each one landed in my tests.

Kling 4.0

Best value for cinematic work

Cheaper than Seedance and the most reliable at holding a real performance while changing who's on screen. The one I'd build a cinematic pipeline on.

Wan 3.0

Animations and hyper motion

About 4.5x cheaper than Seedance 2.5 at matched settings, and it punches way above that price on motion design, stylised animation and hyper motion shots.

Seedance 2.5

Overall quality winner

The clear winner on pure image quality. The premium price is justified when the piece is a film or anything that demands high production value.

Kling 4.0Fast action, tracking shot. A horse at full gallop through a tree line, foreground trunks sliding past, then the camera picks the rider up side on. Legs, dust and the rider's face all stay stable. Fast scenes are where older Kling versions used to fall apart.

How to read this guide

This post focuses on Kling 4.0 because it does the most for the money: long takes, fast movement and a full edit mode in one model. Parts 2 and 3 cover the specs and the cost math. Parts 4 to 6 are the full workflow, the worked prompt, and why each part of the prompt is there. Part 7 is the test order and fixes. Part 8 tells you when to pick Wan or Seedance instead.

Everything here runs inside Higgsfield. All three models live in the same workspace on the same credits, so you can switch models without changing tools: higgsfield.ai/ai/video

Part 2 · Kling 4.0 spec card

What you're working with on Higgsfield right now.

Modes

Text, Image, Reference, Video edit

Length

3 to 30 sec, one pass, no stitching

Resolution

720p / 1080p, native 4K announced as coming

Aspect ratios

9:16, 16:9, 1:1, 21:9, Auto

Keyframes

Up to 10, pin poses or beats at any second

References

10 img + 5 video

Video edit input

Up to 5 clips, 1 main clip, 30 sec combined

Audio

Stereo, 9+ languages

Prompt length

Up to 8,000 tokens

The four modes, and which one this guide uses

Mode	What it does	Use it for
Text-to-video	Prompt only.	Ideas, b-roll, fast tests
Image-to-video	Start / end frame plus keyframes in between.	Controlled product shots
Reference-to-video	Up to 10 images + 5 videos to lock characters, products and motion.	Consistent characters across a series
Video edit	Change the character, prop, background, angle or style inside an existing clip.	Changing cast in a shot: this guide

Kling 4.0Stable in fast motion. Low tracking shot on a galloping horse, sand kicking up, hard backlight, then the rider's face comes into frame and holds. This is the more stable in fast scenes upgrade over Kling 3.0 in practice.

Part 3 · The cost math

Cinematic shots are where the budget goes. This is how you know what a sequence actually costs before you render it.

Kling 4.0	5 sec	10 sec	30 sec	Per second
720p	25 cr	50 cr	150 cr	5 cr
1080p	37.5 cr	75 cr	225 cr	7.5 cr

Kling 4.0 vs Seedance 2.5, per second

Kling · 720p

5 cr

Seedance · 720p

7 cr

Kling · 1080p

7.5 cr

Seedance · 1080p

16 cr

At 1080p, Kling 4.0 costs less than half of Seedance 2.5 per second. That's the gap that matters when you're building a sequence of 20, 30, 50 shots.

Real example: one sequence, 30 shots

30 x 10 sec shots at 1080p	Credits	Approx. cost on Plus
Kling 4.0	2,250	about $110
Seedance 2.5	4,800	about $235

Approx. dollar costs use the Higgsfield Plus plan ($49 / 1,000 credits, about $0.049 per credit). Seedance figures are based on 56 cr (720p) and 128 cr (1080p) per 8 sec clip. Wan 3.0 came in about 4.5x cheaper than Seedance 2.5 at matched settings. Prices change, so always check the number on the Generate button before you render.

Draft cheap, finish sharp. Lock the motion at 720p, 5 sec (25 cr), then render the keeper at 1080p. You'll only pay full price on clips you already know work.

Kling 4.0Three beats, one generation. Miniature building with neon icons, push in through the lit window, the man at the desk, close up. No stitching. At Kling 4.0 rates, 10 seconds like this is 50 cr at 720p or 75 cr at 1080p.

Part 4 · The workflow: one performance, any cast

The idea: record one strong performance, then use Kling 4.0's video edit mode to change who is on screen while keeping the movement, timing and gestures. One shoot, a whole cast.

1
Record the base performance
Shoot the performance you want to keep, 10 to 15 sec. Clear gestures, a real expression, hands visible.
- One continuous take, no cuts. Cuts break the motion lock.
- Keep the subject fully in frame with space around the body.
- Even light, clean background. Nothing branded in shot.
2
Create the new cast
Generate 3 to 5 character images in Higgsfield's image tools. Each one should match the base clip's framing: same crop, same angle, similar pose, hands visible, neutral background. The closer the framing, the cleaner the swap.
3
Set up Kling 4.0
Higgsfield, Video, Kling 4.0, Edit video. Upload the base clip as the main clip, add the character image as a reference.
- Aspect ratio: match the base clip
- Resolution: 720p for drafts, 1080p for finals
- Duration: match your base clip
4
Write the swap prompt
Use the worked prompt in Part 5. Tag the character and any hero prop so the model knows which is which.
5
Generate and check it against the original
Play the result next to the source clip. Check gesture timing, hands, lip sync and props. If something is off, go to the fixes in Part 7.
6
Scale it
Lock the prompt and settings. Swap only the character reference. Every new image is a new cast member in the same shot.
7
Finish
Grade and cut in DaVinci Resolve or CapCut. For the reveal, build a synced before/after split screen: original on the left, swapped version on the right.

Part 5 · The full worked prompt

Mode: Edit video · Main clip: base performance · @image1 = new cast · @image2 = hero prop (optional)

[SWAP]
Replace the person in the video with the person in @image1.

[LOCK THE PERFORMANCE]
Keep the original performance exactly: same head movement,
hand gestures, timing, lip movement and eye line to camera.

[PROP]
Keep the prop from @image2 in hand, same size and position
as the original.

[SCENE]
Same room, same lighting, same camera and lens framing.

[LOOK]
Natural skin texture, no beauty filter, no added camera moves.
Keep the original grade.

Kling 4.0Identity held through a turn. A backlit silhouette turns his head and the face, glasses and hair read as the same person. The turn is where a swapped identity usually breaks, so it's the first thing I check on every result.

Variation prompts

Want to change	Add this line to the prompt
Location	Move the scene to a rain lit street at night. Keep the camera framing.
Wardrobe	Dress the person in a dark coat. Everything else unchanged.
Language	Same delivery, spoken in Spanish, lip sync matched.

Part 6 · Why each block is there

Block	Job	What happens without it
[SWAP]	Tells the model the only thing that should change is the person, and points it to one exact reference.	Kling may restyle the whole frame or blend the two faces.
[LOCK THE PERFORMANCE]	Names the specific motion to keep: head, hands, timing, lips, eye line. This is the motion control.	Gestures get smoothed, timing drifts and the delivery loses the energy that made the take work.
[PROP]	Protects the object the shot is built around.	Props warp, change shape or end up in the wrong hand.
[SCENE]	Holds lighting, camera and lens so the new cast sits in the same world.	Lighting mismatch: the new face looks pasted in.
[LOOK]	Keeps the graded, filmic feel intact.	Plastic skin and a flat grade that reads as generated.

Rules I stick to

Short and literal beats long and poetic. Each block is one job. The model doesn't need adjectives to hold a gesture.
Say what to keep, not just what to change. The LOCK block does more work than the SWAP block.
One reference per job. One image for the person, one for the prop. Don't stack three faces and hope.
Change one thing at a time. If a result misses, change the prompt or the reference or the duration, never all three.

Part 7 · Test order and fixes

Run your tests in this order

#	Test	Settings	Pass when
1	Motion lock	720p · 5 sec · first 5 sec of base clip	Gestures match the source
2	Full length	720p · full clip length	No drift at the end
3	Final render	1080p · full clip length	Hands, props, lips clean
4	Scale	Same as #3, new @image1 only	Same quality per cast member

Common problems

Problem	Fix
Face drifts back to the original	Use a sharper, front facing reference with the same head angle as the base clip.
Hands melt or merge with the prop	Re-shoot the base with hands away from the face and the prop held clearly. Add a prop reference.
Gestures get softer	Strengthen the LOCK block. Name the exact gestures: "points at camera at 3 sec".
Lip sync slips	Keep base clips under 15 sec. Make sure the mouth is clearly visible and well lit in the source.
Swap looks pasted on	Match the reference image's lighting and crop to the base clip before generating.
Background changes	Add "background unchanged" to SCENE.

Want image + motion instead of editing a clip? Higgsfield also has Kling 3.0 Motion Control and Genjutsu: give it a character image and a motion reference video and it transfers the movement. Good for dance and full body moves.

Part 8 · When to use which model

Kling isn't the answer for everything. Here's the call I make after running all three side by side.

Kling 4.0

Cinematic at volume. Motion control: change who is on screen, keep the performance. Less than half Seedance's cost per second at 1080p, so you can shoot far more coverage.