The world reacts to motion
Lesson 11 of 22 in the course “AI Video, No Flops”. Free, with animations and a quiz.
Action and reaction
A person has weight not only because they stand right but also because the world reacts to them. Newton’s third law: the foot pushes the ground back, the ground pushes the body forward, and dust flies from under the sole. A hand pushes a door, and the door swings. A skater pushes off, and the wheels clack. When OpenAI introduced Sora 2 in 2025, it gave exactly this kind of example: if a basketball player missed, older models might “teleport” the ball into the hoop, while the new one bounces it off the backboard. So even model makers see the physics of reaction as the main sign of realism. Your job is to describe not only the motion but also what it changes around it.
Consequences stay
In the first Sora announcement in 2024, OpenAI openly listed the model’s weaknesses. One of them: a person might take a bite out of a cookie, but afterwards the cookie may have no bite mark. It’s a typical error: the model draws the action but forgets its result. Footprints in the snow vanish, wet hair is dry a second later, a crushed cup is smooth again. The fix is to write the result as a separate phrase: “a visible bite mark remains”, “her footprints stay in the snow”. And if the model still loses it, make the next shot already show the result, for example a close-up of the bitten burger.
Fabric by material
For a clothing brand, this is the key slide of the lesson. The model doesn’t know what your hoodie is made of until you say so, and the whole motion depends on the material. Thick cotton fleece is heavy: it drapes and swings slowly and softly. A nylon windbreaker is light: it rustles and flaps in the wind. Silk flows and follows every movement of the body. Raw denim is stiff and holds its creases. Write the material and fabric weight right into the clothing description: “oversized heavyweight cotton hoodie, 400 gsm”. A shopper in the feed feels the thickness of the fabric from how it moves before they read the description.
One wind for everything
Wind is a common source of quiet glitches. The model animates each detail separately, and in one frame hair flies left, a flag flies right, and smoke rises straight up. The viewer may not know what’s wrong, but they’ll feel it’s fake. The rule is simple: a scene has one wind direction, and every light object obeys it: hair, drawstrings, fabric, leaves, smoke, flags. Write it explicitly: “strong wind from the left, her hair and hoodie strings blow to the right”. If you don’t want wind, say that too: “still air”, so nothing flutters for no reason.
Splashes, dust, snow
Particles are a cheap way to sell weight. A step into a puddle throws splashes in the step’s direction and sends ripples across the water. Landing on a dusty floor raises a puff at the feet. Snow and sand keep footprints, and their depth hints at weight. Proportion matters: the harder the impact, the more particles and the farther they fly. If a skater lands from a metre up and not a speck of dust rises under the wheels, the impact feels weightless. If a light step makes water shoot up like a fountain, that’s comedy. Write particles next to the action: “lands, a small puff of dust rises around the wheels”.
Shadows and reflections
A shadow is proof of contact. While a person stands, the shadow starts right at the feet. In a jump the shadow separates from the legs and stays on the ground, and the brain reads the height from that gap. Models sometimes forget the shadow: the person looks pasted over the background like in a bad collage. Reflections work the same way: in a puddle, a shop window or wet asphalt, the reflection has to repeat the motion. The check is simple: pause the video mid-jump and look for the shadow. If the shadow is stuck to the legs in the air, or there’s none at all, redo the shot.
Gravity is the same for everything
Everything that falls, falls with acceleration: slowly in the first fraction of a second, then faster and faster. A thrown cap flies in an arc, a parabola. That’s school physics, and the brain knows it intuitively, so a phone gently floating down or a cap flying in a straight line instantly gives AI away. Models love “pretty” slow motion, so if you don’t want slow-mo, write “real-time speed”. And if you do want it, slow everything down on purpose and all at once: in real slow-mo the person, the hair and the splashes all move slowly, not just one thing. Look at Mira’s shots: the cap flies in an arc, the phone falls faster and faster, and it reads as real.
Check yourself
- For a collab with a burger place, Mira films a bite. A second later the burger is whole again. What should she do?
Correct answer: Add “a visible bite mark remains” or show the bitten burger in the next shot. The model forgets an action’s result unless it’s stated. A separate phrase or shot locks the mark in. - In the rooftop shot, Mira’s hair blows left and the hoodie strings blow right. How to fix it?
Correct answer: Set one wind: “wind from the right, her hair and strings blow to the left”. One explicit wind direction makes every light object in the frame obey it. - Mira tosses a folded hoodie into a box, and it drifts down slowly like a feather. What should she add?
Correct answer: “the folded heavyweight hoodie drops quickly and lands with a soft thud, real-time speed”. Heavy folded fabric falls fast with a dull thud. Speed, material and sound together create weight. - A skater lands from a metre up, but there’s no sound and no dust. What sells the impact best?
Correct answer: “lands, a small puff of dust rises around the wheels, a sharp clack”. Particles and a sharp sound at contact are the world’s main response to an impact. The brain believes the weight instantly.
← Phases of motion and the motion prompt Motion control and the weight test →