Centre of mass and support
Lesson 8 of 22 in the course “AI Video, No Flops”. Free, with animations and a quiz.
Centre of mass over the support
We’re starting the key level of the course: how to make the person on screen have weight. Every body has a centre of mass. In a standing person it’s roughly at the level of the sacrum, just below the navel. The balance rule from biomechanics is simple: if the vertical line from the centre of mass falls inside the base of support (between the feet), the person stands. If it goes past the edge, they fall or take a step to catch themselves. Video models learned from pixels, not physics, so they easily draw people “hanging” in a lean who should have fallen long ago. The viewer’s brain feels it instantly, even if they can’t explain why.
A load changes the pose
Hold a heavy bag in one hand and look in the mirror. Your torso leans the other way on its own, and your free arm moves out for balance: the body keeps the combined centre of mass over the feet. If a person in an AI video carries “heavy” bags while standing perfectly straight, the bags look empty. The good news: models draw these poses fairly well if you ask in words. Don’t write “carries a bag”; write “carries a heavy bag in her right hand, torso leaning left, short steps”. The more specific the load, the more honest the body.
Feet don’t slide
Animators and motion-capture specialists have a term, foot sliding: a foot drifts across the floor even though all the weight rests on it. In real life that doesn’t happen: the support leg stays in place while the body moves over it, and lifts only once the weight has moved to the other leg. It’s the most common AI-walk glitch: the person seems to walk on ice or float over the asphalt. Two things help. First, describe the contact: “heel strikes, the foot stays planted”. Second, don’t ask for a very fast walk and a complex camera move at the same time: the more things move, the more often the model “lets go” of the feet.
Heavy things move differently
Watch any weightlifting video: before the lift the athlete squats, back straight, the bar rises slowly, and at the top they freeze for a moment. Weight shows not in the size of the object but in how the body prepares for it and how slowly it moves it. Light things are picked up with one hand, quickly, without changing the pose. By default a model often makes everything equally light: a 20 kg box flies up like an empty one. So don’t just say “a heavy box”; describe the body’s reaction: she squats, tenses up, lifts slowly, pauses.
Jump and landing
Landing is the best weight test. A real person lands in three phases: in flight the legs are tucked; on contact the toes touch first, then the heel; then the knees bend deeply to absorb the impact while the arms go forward for balance. Look at gymnasts and parkour athletes: the higher the jump, the deeper the squat after it. If the model draws a landing on straight legs, the person looks like a feather or a doll. Write the phases right into the prompt: “lands from the jump, knees bend deeply and absorb the impact”. And if the landing still breaks, an editing trick: cut to another shot at the moment of contact.
Words that add weight
Build yourself a little dictionary. The model doesn’t understand “make it realistic”, but it understands specific physical actions well. Support: “plants her feet shoulder-width apart”. Weight shift: “shifts her weight onto the left leg”. Effort: “bends her knees, lifts the heavy box slowly, arms straining”. Absorbing: “knees absorb the impact”. Contact: “heel strikes, the foot stays planted”. Use English for these: models learned most from English descriptions. One or two of these verbs per shot, and the body starts to have weight.
A 10-second weight check
Before posting, run every shot with people through five questions. Where’s the support, and is the centre of mass over the feet? Do the feet stay put while bearing weight? Do the knees absorb steps and jumps? Does a load change the pose and speed? And most important: don’t watch at normal speed; go frame by frame or at 0.25×. At full speed the brain forgives a lot, but viewers rewatch an ad on loop, and by the third time someone spots the glitch. Mira does this before every drop: five questions, a slow watch, and only then the post.
Check yourself
- Mira’s model walks down the street, but her feet glide as if on ice. What helps first?
Correct answer: Describe contact, “heel strikes, the foot stays planted”, and simplify the camera move. The model keeps the foot planted when contact is spelled out, and it copes better when less is moving in the frame. - Mira carries a box of 30 hoodies, but in the video the box looks empty. What should she add?
Correct answer: The body’s reaction: knees bent, torso leaning back, short slow steps. Weight shows in how the body copes with the load. Describe that, and the box gets heavier. - A skater in the video jumps off the steps and lands on straight legs. What should be fixed?
Correct answer: Add the phases: “lands, knees bend deeply and absorb the impact, arms forward”. Knees absorbing the impact is the main sign of weight in a landing. The model draws it when you mention it. - The hero has a heavy bag on the right shoulder. Which way should the torso lean?
Correct answer: Left, away from the load, so the centre of mass stays over the feet. The body balances a load with a counterweight. Leaning away from the bag looks honest and heavy.