Kling 3.0 Motion Control Guide: Driving Video, Orientation Modes and Limits

All guides · Kling 3.0 Motion in Fuser

Quick answer: Kling 3.0 Motion Control takes two inputs, a still image of your character and a video of someone performing the motion, and returns a clip of your character doing that motion in the image's setting. Use one person, whole body or upper body visible in both inputs, framed the same way, and a driving clip of 3 to 30 seconds with no cuts. Leave Character Orientation on Video for full-body or complex movement (driving clip up to 30 s); switch to Image when the character should keep the pose and facing of your still (driving clip up to 10 s) (Kling motion control guide, Kling motion control API reference). If the character has to touch a prop, put the prop where the performer's hands are. The prompt is optional. For writing text-to-video and image-to-video prompts, see the Kling 3.0 prompt guide; this page covers only motion transfer.

What motion control does, and what it doesn't

A normal image-to-video model invents motion from a prompt. Motion control copies it: body pose, gestures, head turns and expressions come from the driving video, while the face, clothes, background and light come from your image. That makes it the tool for a specific dance, a product demo gesture, or an actor's timing that you already have on video.

It copies motion literally. It does not work out what the motion was for. In our first test below, the driving clip is a potter shaping a vase on a wheel and the character image is a baker with a ball of dough. The baker's hands repeat the potter's shaping gestures in the air above the dough and never touch it, even though the prompt asked for kneading. A matching action is not enough on its own either: the objects have to sit where the performer's hands are (see the layout test below).

The two inputs

Character image. One clearly visible person with body proportions that read clearly and nothing covering them, filling more than 5% of the frame. The image also sets the background and props, so leave space around the character to move (Kling motion control guide). On the API, the image's aspect ratio must be between 0.4 and 2.5 (Kling motion control API reference).

Driving video. A realistic person with the whole body or upper body and head in view, 3 to 30 seconds, up to 100 MB (Kling motion control API reference). Kling's guidance for the clip:

• One person. With two or more people, only the largest figure's motion is used.

• One continuous shot. Avoid cuts and shot changes in the driving clip.

• Moderate speed. Very fast motion gives poor results.

• Matching framing. Full body with full body, half body with half body.

All four points are from the Kling motion control guide. The output length follows the driving clip, so trim it to the moment you want before you run.

Video or Image orientation

Character Orientation is the one setting that changes the result most.

• Video (the default): the character's facing, pose and expressions follow the driving clip. It is better for complex and full-body motion, and accepts driving clips up to 30 seconds.

• Image: the character's movements still follow the driving clip, but the character keeps the facing of your still. Kling describes it as the mode that supports camera movement, and it caps the driving clip at 10 seconds.

Those descriptions are Kling's (Kling motion control guide, Kling motion control API reference). Here is what they looked like on the same inputs:

• Video orientation turned the baker three-quarters to her right, the way the potter faces her wheel, and her hands worked the air to the right of the dough, roughly where the potter's vase stands. At the end the baker looks up and speaks to camera, as the potter does.

• Image orientation kept the baker square to camera behind the dough, close to the facing of the still, and ran the same gestures from that facing, hands in front of her chest. It returned 3.7 seconds from a 4-second driving clip; Video returned the full 4.0.

If the still already shows the angle you want, try Image first. If the motion only makes sense from the performer's angle, as with a turn, a walk or a dance, use Video.

Match the layout, not just the action

Because the first test mixed two actions, we ran a second one on the Pro tier with a matching action: the same potter clip, this time with an older potter at a wheel as the character. We ran it twice with two different character images and changed nothing else.

• Image A is a fresh FLUX.2 render: a potter at a wheel in a different studio, sitting further left, with the vase further to the right than in the clip. In the output his hands make the shaping movements in the air to the left of the vase and never reach it, the same failure as the baker's.

• Image B is the driving clip's first frame, edited with FLUX.2 to replace the woman with the same kind of older potter, so the wheel and vase sit where hers do. In the output his hands close around the vase and follow its shape, then he looks up and smiles to camera as she does.

The model places the hands where the performer's hands were in the frame. It does not move them to meet an object. When the character has to touch something, the easiest way to get it right is to build the character image from the driving clip's first frame, as we did for image B, or at least to compose it with the same camera angle and object positions. Neither Pro output needed a prompt.

Sound

Keep Original Sound carries the driving clip's audio onto the output. It is on by default. In our run both outputs kept the potter's workshop sound and her spoken line at the same level as the source clip, so the baker appears to say her words. Turn it off if you plan to add a voice or sound later, for example with a lip-sync model or a sound effects model.

Standard or Pro

Both tiers take the same inputs and settings. Fuser defaults to Standard. Pro costs a third more per second of output than Standard (Kling motion control guide). In our runs the Standard outputs came back at 1280 × 720 and the Pro outputs at 1920 × 1088, from 720p driving clips in both cases. Get the driving clip, the layout and the orientation right on Standard, then rerun the keeper on Pro.

The prompt

The prompt is optional (up to 2,500 characters). It does not override the driving clip: in our run, "works the dough with both hands" did not change the transferred hand movement. Keep motion in the driving clip and use the prompt, if at all, for details the image and video don't already fix.

Kling's API also accepts a facial reference ("element") to hold identity in Video orientation only. The Kling 3.0 Motion node in Fuser does not expose it, so for identity, start from a strong, well-lit character image. See consistent characters across images and video for how to make one.

Build it in Fuser

The canvas at the top is the first test. The character image is a FLUX.2 render; the driving clip is a Veo 3.1 clip from our Seedance vs Kling vs Veo test, trimmed to four seconds. Both connect to two Kling 3.0 Motion nodes set to different orientations, so the two results sit side by side.

1. Add or generate the character image and connect it to the node's Image input.

2. Upload a performance video, or generate one, and connect it to Reference Video.

3. Pick the orientation, leave the model on Standard and run.

4. Duplicate the node to try the other orientation on the same inputs.

5. Send the result on to an upscaler or a lip-sync step.

For performance transfer focused on faces and dialogue, compare Runway Act-Two. For talking-head presenters, see the best AI avatar video generators. When you only need a generated shot rather than a copied performance, Kling 3.0 Video is the right node.