The Physics AI Video Still Gets Wrong

Ask for a boulder rolling down a hill and there’s a real chance it bounces like a beach ball. Nothing about the scene looks obviously wrong at first glance, which is exactly why AI video’s physics problem is easy to miss until it’s already happened.

Liquids are the hardest material there is

Splashing, pouring, waves breaking: all three involve fast, chaotic detail that ordinary footage rarely captures cleanly enough to learn from. The result tends to move like something thicker than water, closer to gel, with a stream that curves a little too smoothly and a splash that resolves too fast or not at all. Liquid shots hold up best kept short, two or three seconds, cut right after the moment of contact rather than asked to carry a full pour from start to settle.

Cloth and hair guess, they don’t simulate

Nothing underneath a video model is running an actual physics engine. Fabric and hair are visual patterns the model has seen move a certain way in open space, not objects with mass and tension calculated frame by frame. That approximation holds up fine for a scarf drifting in open wind. It struggles the moment the shot depends on a precise mechanical interaction, a scarf catching on a branch and pulling taut, a hand pinning down a corner of fabric against the wind. If a shot’s entire point rests on that kind of contact, it’s worth simplifying the setup rather than trusting the model to resolve it correctly.

Weight is a suggestion

A model reads mass mostly from shape and texture, not from anything resembling physical weight, so a boulder can end up moving with the same bounce as a beach ball if the prompt never says otherwise. Naming the material and the resulting motion directly, drags, thuds, sinks into the ground, gives the model the physical cue it has no other way to infer, instead of leaving weight up to a guess based on what a round gray object usually does in the training data.

Fire acts like a light source more than a flame

Real fire is chaotic in a specific way, it flickers, ducks, and stretches based on airflow that’s genuinely hard to predict even in person. Generated fire often settles into something closer to a glowing, gently pulsing light effect than actual combustion, with flame shapes that repeat a little too smoothly and smoke that drifts without much relationship to anything resembling wind. It’s a hard one to fully fix with wording alone, but naming a specific air condition, still air, a strong gust from the left, gives the model at least one physical cause to react to instead of animating flame in a vacuum.

Small, fast, and repeating breaks first

A bouncing ball, a spinning coin, a ticking second hand: anything meant to repeat at a steady rhythm is a rough match for a model generating frame by frame without a true sense of periodic motion. The rhythm tends to drift a little, a bounce that lands slightly early, a spin that slows without warning, since nothing underneath is enforcing a fixed cycle length. These shots hold up better trimmed to a single repetition, one bounce, one full spin, rather than asked to hold a steady rhythm for several cycles in a row.