Motion control and the weight test
Lesson 12 of 22 in the course “AI Video, No Flops”. Free, with animations and a quiz.
Motion from a video
Everything we described in words in this level can be given to the model directly, as live video. That’s motion control: you give three inputs. The reference video sets the motion and timing. The image sets the character, meaning who moves. The prompt describes the scene, light and style. According to Kling’s guide, the prompt in this mode shouldn’t describe the action again: it’s already in the video, and the text would fight it. Why it’s powerful: a real person in the reference already moves with the right weight, anticipation and arcs, and the model transfers that physics to your character. Complex choreography you can’t put into words gets solved in a single take.
A good reference
A motion-control result depends on the reference more than anything else, so film it by the rules from Kling’s guide. 3–30 seconds, one continuous take: cuts and camera moves confuse the system. One person in frame; with several, the model picks the largest. The whole body and head, with room at the edges so the arms aren’t cropped. Moderate-speed, broad motion without running across the frame: very fast and complex moves transfer poorly. A plain background and good contrast so arms and legs are easy to track. Mira films references on her phone on a tripod against a white wall: it’s free and works better than any pretty clip.
Framing must match
The most common beginner mistake: a pretty close-up portrait plus a full-body dance. The model has nothing to fill the body that isn’t in the image, so it starts inventing arms and legs, and the face drifts. Kling’s guide warns directly: if the video is full-body, the image must be full-body too. Same with the angle: if the person in the reference stands at three-quarters and the image is straight-on, you get jitter and proportions drifting. Before uploading, put the image and the video’s first frame side by side and compare: same shot size, same angle, the whole body visible.
Two orientation modes
In Kling, motion control has two modes. “Matches video” is the default: motion, expressions, camera and body orientation follow the reference, and the prompt can only adjust details. It’s the best choice when you filmed the reference yourself and like all of it. “Matches image” takes motion and expressions from the video, but the pose and body orientation stay as in the image, and you set the camera movement in the prompt. That’s handy when you have a perfect character frame, say Mira at three-quarters by a shop window, and you don’t want to break it. Try both modes on one draft: the difference is obvious.
Open source: Wan-Animate-2
If you want full control and don’t want to pay for every generation, there’s an open path. Wan-Animate-2 came out in August 2026 under the Apache-2.0 licence, which allows commercial use. The model animates a character from an image using motion from a video: a pose skeleton is extracted from the reference, and the character moves along it. It runs in ComfyUI, a free program where you build the pipeline from node blocks like a construction kit. You need a powerful graphics card or a cloud server with a GPU. The downside is the barrier to entry: setup takes an evening, not a minute. The upside: no credits, your own hardware, and every parameter under your control.
Whose motion is it?
Motion control transfers any video so easily that it’s tempting to grab another creator’s viral dance and “put it on” your model. For an ad, that’s a bad idea. Someone else’s video is someone else’s content, and being publicly available doesn’t give you the right to use it in a commercial video. The creator can file a complaint with the platform, and for a brand that’s a hit to its reputation. Safe options: film yourself or a friend on a tripod, use licensed stock footage, or agree with the creator in writing. We’ll talk about other people’s faces and voices in detail in the publishing level.
The weight test
The level finale is a checklist that pulls everything together. Before posting, run every shot with people through it. The centre of mass is over the support, and a load changes the pose. Feet don’t slide, knees absorb steps and jumps. There’s anticipation, arcs and follow-through in fabric and hair. The world responds: dust, sound, shadows, lasting marks. And all of it is checked frame by frame at 0.25×, not by eye. If a shot passes all five, viewers will say “that’s filmed”, even if they can’t explain why. If not, you know exactly which one thing to fix in the next iteration. Now let’s lock it in with scenarios.
Check yourself
- For motion control Mira uploaded a close-up selfie and a full-body dance video. The arms and legs came out weird and the face drifts. Why?
Correct answer: The framing doesn’t match: a full-body dance needs a full-body image. The model has nothing to fill the missing body with, so it invents it. Kling’s guide explicitly requires matching framing. - Mira filmed the reference handheld, with two cuts. The motion transfer jitters. What should she do?
Correct answer: Reshoot in one continuous take on a tripod. Cuts and camera motion confuse the system. A steady continuous take is the main reference requirement. - In the motion-control prompt Mira described the dance again in detail. The motion got worse. What should the prompt say?
Correct answer: The scene, light and style, and leave the action to the video. The motion is already in the reference, and text about the action fights it. The prompt handles the world around: place, light, clothing, style. - Mira found another creator’s viral dance and wants to transfer it to her model for the drop ad. What’s the right move?
Correct answer: Film her own reference or get the creator’s written agreement. Someone else’s video is someone else’s content. A commercial video needs your own material or explicit permission.