Kling 2.6 setup guide

Kling 2.6 Motion Control: Set Up the Inputs Before You Generate

For Kling 2.6 Motion Control, start with one continuous action video and one character image whose crop, body scale, and facing direction make the action physically possible. Then decide whether orientation should follow the video or the image. Do that before adding scene detail to the prompt.

The feature has two jobs that are easy to mix together: the action video provides movement, while the character image provides the person or character who performs it. A useful first render makes those jobs easy to inspect. It is not meant to be the final styled scene.

Use this input pairing test

Look at the first clear pose in the action video, then look at the character image. Answer four questions before uploading.

Can the image contain the action? A waist-up image is a weak match for a large step, turn, or arm sweep that needs visible hips and legs. Start with a full-body image for full-body movement, and a half-body image for half-body movement. The goal is to give the model visible room for the parts that need to move.

Does the action start from a plausible direction? A left-facing performer in the video and a strongly right-facing character image may create a conflict before the motion begins. Do not hide that conflict under a detailed prompt. Choose a closer match or change the intended orientation deliberately.

Is one performer visually dominant? Kling's guide says a reference containing more than one person uses the person with the largest on-screen presence. If two people occupy similar space, no element may be selected. A first test should make the intended performer obvious in every frame.

Is the action one unbroken shot? Cuts, shot changes, and camera movement make it harder to preserve a continuous action. Use a clip where the performer remains visible, rather than a montage that only looks energetic when watched at normal speed.

Choose orientation as a diagnosis tool

Kling 2.6 offers an orientation choice that either matches the character to the video or matches the character to the image. The first choice is the straightforward option when the motion reference already has the desired direction. The second can be useful when the character image must keep its own direction and the scene needs camera movement controlled through the prompt.

Make two deliberately small tests when direction is uncertain. In the first, let orientation match the video. In the second, let it match the image. Keep the same action video, image, and simple scene description. Comparing those two runs tells you whether the problem is the input pair or the orientation choice. Changing every variable at once tells you almost nothing.

Keep the first action easy to evaluate

Official guidance recommends a single character, a wide but moderate range of movement, limited displacement, and a steady speed. That advice is more useful when treated as a test design rule: begin with a motion whose first beat, middle, and finish are all visible. Avoid a very fast action until the simpler version has shown that the crop and direction work.

For uploaded action videos, Kling's guide specifies a duration of 3 to 30 seconds, a short edge of at least 340 pixels, and a long edge no greater than 3850 pixels. The guide also notes that a complex or fast action can produce a shorter result when only a valid continuous segment can be extracted. Treat duration as something to verify in a trial, not as proof that every frame of the reference will survive.

Separate the prompt from the performance

Let the action video carry timing, gesture, pace, and expression changes. Use the prompt for the setting, lighting, atmosphere, and other scene details. When the prompt re-describes the same movement with different wording, an unsuccessful render gives no clear clue about which input should change.

A clean first prompt can be short: name the setting and visual treatment, then stop. If the basic motion works, add one visual layer at a time. This makes a later failure traceable to a background change, a camera instruction, or a more demanding action rather than to a long list of simultaneous changes.

Review the result in a fixed order

Do not judge a render only by its best frame. First check the direction change. Next watch the face through expression changes. Then inspect hands. Finally, inspect any moment when the performer touches or carries an object. These are separate tests: a convincing torso does not establish that facial detail, fingers, and object contact are equally stable.

If the result breaks, change the earliest failed input gate instead of restarting from a new idea. A crop mismatch calls for a different character image. A discontinuous action calls for a different video. A direction conflict calls for a different orientation test. That small decision rule makes Kling 2.6 Motion Control easier to repeat—and easier to improve—than a sequence of unrelated rerolls.