Kling motion control: how to pick the photo and the clip
Practical requirements for the character photo and the motion reference video: framing, character orientation, prompting, and fixes for common artifacts.
Motion control works differently from regular video generation. There is no prompt that decides the outcome: you provide a photo of a character and a clip with movement, and the model transfers one onto the other. That is why almost every failed attempt comes down to unsuitable source material, not to the model.
Below are the requirements that actually affect the result: Kling's official limits, advice from guides, and what we verified in our own runs of motion control.
In short
- The face in the photo is large and sharp, and the hands are visible.
- There is room around the character at the edges of the frame.
- The clip shows one person in one continuous take, with a static camera.
- The movement is moderately paced, without sudden jerks.
- The prompt does not describe the movement — only the setting and the light.
- Draft at 720p on a short fragment, render the final at 1080p.
The character photo
The model takes everything from the photo except the movement: the face, clothing, background, proportions and framing. So the quality of the photo sets the ceiling for the quality of the result.
The face should be large and sharp. The better the eyes and lips are visible, the more accurately facial expressions transfer. A small, blurry or overexposed face is the main reason a "different person" shows up in the video.
The hands should be visible. If the hands are in pockets or behind the back in the photo, but the character gestures in the clip, the model has to draw the hands from scratch. That is where extra and twisted fingers come from.
Leave some air at the edges. Kling requires the character to take up more than 5% of the frame, but the opposite extreme hurts too: if the figure touches the borders and the clip has wide swings or jumps, limbs will get cut off.
One person. If there are two people in the photo, the model takes whoever takes up more space. If they take up about the same area, it takes no one.
File size. The official limit is 3850 px on the long side. A shot from a modern phone exceeds that, so we downscale photos automatically: upload the original, no preparation needed.
The motion clip
The clip is not "inspiration" — it is the literal source of the trajectory. The model reads the pose frame by frame, so a messy source breaks the result more than a bad prompt does.
One person and one continuous take. Cuts, angle changes and camera zooms are not supported. With several people in the frame, the movement comes from whoever is larger — and that is not always the person you meant.
The camera stays still. Handheld footage and camera moves cause a drifting background and "rubbery" limbs: the model cannot separate the person's movement from the movement of the frame.
Moderate pace. Fast choreography and sudden jerks cause morphing. Slow, clean movement transfers much more reliably than complex movement.
The background should contrast with the figure. If the person blends into the background, the model loses limbs.
Duration is 3 to 30 seconds, with an important caveat: the model only extracts continuous stretches of movement at least 3 seconds long. If the person leaves the frame for a second or is blocked by an object in your fragment, the result will be shorter than you ordered — while you pay for the length of the selected fragment. So pick the fragment carefully with the slider: exactly what you select goes into the generation.
Framing: the most persistent myth
Motion control guides keep repeating a rule: the photo must be full-length, or the proportions will "drift". Kling's official documentation puts it more softly — it advises against pairing a full-body reference with a waist-up photo.
We tested it. A waist-up portrait plus a dance reference shot full-length: the model kept the photo's framing and produced a waist-up video. It did not invent legs, and the face, clothing and background stayed in place.
The rule works the other way around from what people think: the result follows the framing of the photo, not of the clip. A waist-up shot gives a waist-up video, and that is fine.
Common sense still applies, though: if the whole point of the movement is in the legs — a walk, a dance with steps — then in a waist-up video you simply won't see what it was all about. Full-length is needed not because something breaks otherwise, but because otherwise the movement won't be visible.
The same goes for aspect ratio: the photo sets the format of the result. A vertical photo gives a vertical video, even if the clip is horizontal.
Character orientation
This is the only setting that really changes how the model behaves, and it is the one people choose wrong most often.
Follow the video — the character repeats both the movement and the body turns from the clip. Use it when the movement includes a change of angle: a dance with spins, a walk, a turn toward the camera. Fragment length is up to 30 seconds.
Follow the photo — the character keeps the orientation of the original photo and moves within it. Use it for portraits and wherever keeping the face matters most. The fragment is limited to 10 seconds.
A practical rule: if the face "drifts" in the result, try "Follow the photo". If the movement transfers weakly and the character looks stiff, try "Follow the video".
Prompting: above all, don't describe the movement
The prompt is optional here, and that is not a formality: the model does fine without it.
The most common mistake is describing choreography in the prompt. "Dances energetically, spins and jumps" doesn't help — it competes with the clip for control. The movement is already set; there is no need to repeat it in words.
What is actually worth writing: the location and background, the light, the mood, the visual style. A short formula works — scene, light, style. Three or four points are better than a paragraph.
The prompt is a handy way to reconcile a conflict between the character and the movement. If the photo shows an evening gown and the clip shows acrobatics, add that the dress moves with her — otherwise the model will choose between "keep the clothing" and "repeat the movement" on its own.
What breaks and how to fix it
| What you see | Cause | What to do |
|---|---|---|
| The face drifted, it's not your person | The face is too small or not sharp in the photo | Crop closer, use a shot with a clear face, try "Follow the photo" |
| Extra fingers, broken hands | Hands not visible in the photo or blocked in the clip | A photo with visible hands; a clip where the hands don't pass in front of the face |
| Limbs stretch and bend | Movement too fast, shaky camera in the reference | A slow fragment, footage shot on a tripod |
| The background drifts and shifts | Camera movement in the clip | A clip with a static camera; anchor the background in the prompt |
| The video is shorter than the fragment | The model found less continuous movement | Pick a fragment where the person never leaves the frame |
| Hands and feet are cut off | The figure touches the edges of the photo | Crop with room at the edges |
How not to waste attempts
Motion control rarely comes out perfect on the first try, and that is normal for every model of this kind. Here is how to save:
- Test the pairing on a short fragment. Five seconds at 720p tell you everything you need to know about whether this photo and this clip work together.
- Render the final at 1080p, with a pairing you have already tested.
- Change one thing at a time. If you change the photo, the fragment and the prompt at once, you won't know what actually helped.
- Start from a ready-made motion template. The clips there are already checked for suitability, so only one variable is left — your photo.
Limits
| Parameter | Value |
|---|---|
| Fragment length | 3–30 seconds |
| With "Follow the photo" | up to 10 seconds |
| Minimum continuous movement | 3 seconds |
| Photo | up to 3850 px on the long side, downscaled automatically |
| Video | up to 150 MB, trimmed with a slider after upload |
| Output quality | 720p or 1080p |
| Prompt | optional |
The audio from the clip stays in the result unchanged — a dance without music is usually not what you want. It also means that the rights to the music in the uploaded file remain with whoever uploaded it.