Kling 2.6 Motion Control: Prompt Patterns That Actually Move the Right Parts

A few days into testing Kling 2.6, I kept tripping over the same snag: I’d ask for a hand wave and the whole shoulder would drift, or I’d ask the camera to push in and the subject would lurch forward like it had been pulled on a string. Not dramatic failures, just that low-grade “move the wrong part” problem that slowly eats trust.

So I set aside an afternoon, made coffee, and tried to understand what the model actually listens to when it decides what to move. What follows isn’t a review. It’s the notes I wish I had on day one: how I nudged Kling 2.6 toward reliable motion control without turning every prompt into a paragraph.

Why “move the wrong part” happens

Kling 2.6 is good at plausible motion. It’s less good at surgical motion. When I asked it to “raise her right hand,” I learned the model doesn’t only parse the words: it also leans on learned priors about how bodies usually move and what the camera usually does. Trust me, that’s when the fun really starts. Here are the patterns I kept seeing:

Ambiguous referents: If there are two plausible actors (subject and camera), Kling sometimes assigns motion to the camera because it keeps scenes stable that way. “Push in” can become subject walking forward.
Occlusion + symmetry: If the right hand is partly hidden or similar to the left in pose, it’ll mirror or swap sides. Side lighting and framing help more than wording here.
Temporal mush: Without beats or duration, movement smears across the whole clip. That’s when you get perpetual micro-jitter or a subject that can’t hold still.

None of this is a bug. It’s the model hedging for realism. The fix, for me, was to talk to Kling the way a motion director talks to a small crew: who moves, which part, how far, how fast, and what stays locked.

Motion tokens checklist

Kling doesn’t require special syntax, but giving myself a tiny “motion token” template kept me honest. I wrote them inline, in plain language, like labels inside the prompt. Here’s the thing — it made a huge difference.

My token gist looks like this:

Actor: who moves (subject, left hand, camera, background light). One actor per token.
Action: the verb, kept small when precision matters (tilt, glance, raise, rotate slightly).
Side/part: exact limb or object (right wrist, left eyebrow, index finger, camera yaw).
Magnitude: small, medium, big, or degrees/counts when needed (15°, 2 steps, 3 frames).
Duration + timing: when it happens and how long (beat 1–2 sec, hold 1 sec, loop).
Constraint: what stays still (torso locked, camera locked, face neutral).
Negative: what must not move (no shoulder shift, no zoom, hair stays calm).

I usually wrap these in brackets to keep them scannable to me, again, for me, not for Kling:

[actor: right hand] [action: wave] [magnitude: small] [duration: 1s] [constraint: shoulder locked] [negative: no camera movement]

Pattern 1: limb control

The limb problem is the classic one. I wanted a subtle hand wave without a full upper-body sway. My first attempts failed because I asked for “a gentle wave.” Kling heard “wave” and animated the whole chain.

What worked better:

Name the smallest moving joint first. “Right wrist flick, fingers lead, shoulder stays neutral.”
Add a counter-constraint. “Torso anchored to chair. No hip sway.”
Give it a beat. “Wave occurs at 0:02–0:03, then rest.”
Keep symmetry in check. “Left hand stays on table, still.”

A real prompt slice that produced clean motion:

“Subject sits at a wooden desk. Calm light. Right wrist makes a small side-to-side wave, fingers doing most of the motion. Shoulder and elbow remain steady. This motion happens once between 2–3 seconds, then the hand returns to resting on the desk. Left hand stays still. No camera move.”

When it matters to be precise, I set a visual anchor: “Right wrist rests next to the coffee mug: only fingers move.” Anchors reduce the model’s uncertainty about where joints live.

Pattern 2: camera

Half of my “wrong part moved” clips were really camera–subject swaps. I asked for a “gentle push-in” and the model made the subject lean forward. Reasonable, but not what I wanted.

How I separated the two:

Declare a camera rig. “Camera on tripod, subtle 5% zoom-in over 2 seconds.”
Freeze the subject. “Subject remains completely still, breathing only.”
Avoid verbs that sound like body actions. I say “zoom-in” or “dolly-in,” not “move closer.”
Use numbers. “5% zoom, linear, 2s, center-framed.”

A prompt slice that behaved:

“Locked-off shot. Camera performs a slow, linear 5% zoom-in from 0–2 seconds. Subject does not lean or step. No parallax change. Background remains fixed.”

If I want parallax (a dolly), I call it: “camera dolly-in, shallow parallax on bookshelf, subject locked to frame center.” That one line stops the model from faking the move with a subject lean.

One more note: Kling 2.6 sometimes adds handheld wobble for ‘cinema.’ If you don’t want that, say it: “no handheld shake, no micro-jitter.”

Pattern 3: micro motion

Micro motion is where Kling 2.6 surprised me in a good way. Small, steady changes tend to look natural — eye saccades, breathing, fabric drift — if you keep verbs gentle and durations short.

What worked:

Eyes: “two small eye saccades left→center in the first second, eyelids soft, head still.”
Breathing: “subtle chest rise 4–6 mm over 3 seconds, then fall, loop once.”
Fabric/hair: “barely-there breeze, hair tips move, roots steady: jacket hem flutters for 0.5s.”
Props: “steam curls from mug, thin, intermittent, not turbulent.”

What didn’t:

Stacking many micros at once. It turns into noise.
Using big verbs with micro targets (“turn head slightly” often becomes a full nod). I use “angle,” “tilt,” “settle,” “drift.”

When clips got too alive, I added a ceiling: “background remains motionless: only [X] moves.” That single line reduced accidental wiggles across the frame.

Negative prompts

Negative prompts are less about banishing artifacts and more about allocating motion budget. I got better results when I used negatives to freeze specific chains.

Useful negatives I leaned on:

“No shoulder involvement.” Forces the model to localize hand motion.
“No camera movement of any kind.” Kills the default wobble.
“Hair stays calm.” Stops the model from “cinematizing” scenes with wind.
“No facial expression change.” Keeps micro motions from drifting into smiles.
“No zoom or dolly.” Avoids the subject/camera swap.

And a small caution: long negative lists can backfire. Two or three targeted negatives beat a laundry list.

For consistency while testing these prompts, we run our Kling experiments on WaveSpeed (our internal GPU environment) to iterate and compare runs without setup drift.

10 ready prompts

These are short, working starters I used. Tweak subjects and styles, but keep the motion bits intact.

1. Clean wrist wave, torso quiet

“A seated person at a wooden desk, daylight.
[actor: right wrist] [action: small side-to-side wave] [magnitude: small]
[timing: 2.0–3.0s, once] [constraint: shoulder and elbow steady]
[negative: no camera movement, no left-hand motion].“

2. Subtle camera push-in, subject frozen

“Portrait framing, neutral background.
[actor: camera] [action: zoom-in] [magnitude: 5%] [duration: 0–2s, linear]
[constraint: subject perfectly still, only breathing]
[negative: no dolly, no handheld shake].“

3. Eye saccades without head movement

“Close-up face, soft light.
[actor: eyes] [action: two quick saccades left→center]
[timing: within first second]
[constraint: head locked, expression neutral]
[negative: no smile, no blink during saccades].“

4. Turn page with finger, not whole arm

“Overhead shot of an open notebook.
[actor: right index finger] [action: slide page corner and flip]
[magnitude: single page] [timing: 1.5–2.5s]
[constraint: wrist minimal, forearm anchored to table]
[negative: no camera move].“

5. Head tilt acknowledgment, micro only

“Medium shot.
[actor: head] [action: small tilt right then return] [magnitude: 5–7°]
[timing: 1.0–1.4s]
[constraint: shoulders level]
[negative: no smile, no eye shift].“

6. Dolly-right with parallax, subject anchored

“Studio scene.
[actor: camera] [action: dolly-right] [magnitude: small, 20 cm]
[duration: 0–2s, smooth]
[constraint: subject remains center-framed, background parallax visible]
[negative: no zoom].“

7. Fabric micro flutter only

“Portrait with light breeze.
[actor: jacket hem] [action: brief flutter]
[timing: 0.8–1.2s]
[constraint: hair and foliage still]
[negative: no global wind].“

8. Keyboard typing with isolated fingers

“Desk scene, laptop.
[actor: fingers both hands] [action: type softly]
[timing: intermittent bursts]
[constraint: wrists planted, elbows steady, torso still]
[negative: no camera shake].”

9. Cup lift without shoulder hike

“Cafe table.
[actor: right hand] [action: lift ceramic cup, sip, replace]
[timing: 1.0–2.5s]
[constraint: shoulder and neck quiet, minimal elbow]
[negative: no camera move, no steam surge].“

10. Background motion while subject holds

“Street portrait.
[actor: background pedestrians] [action: slow pass left→right]
[constraint: subject motionless, eyes at camera]
[negative: no subject sway, no zoom].”

If you need to scale precision, stack one more constraint rather than another action. It keeps the clip readable, to you and to the model. No judgment here — we’ve all been there.

Alright, your turn — don’t leave me hanging: What’s the most ridiculous “open weights” license surprise you’ve hit? Drop your story in the comments!