The physics test: what video models still break, and how to check in ten seconds

Liquids and granular materials

Watch the leading edge of a pour, not the splash. A real stream narrows as it accelerates and breaks into droplets at a distance you can estimate from the height it fell. Generated streams tend to hold a constant width the whole way down, because width is a texture and narrowing is a consequence of something the model is not computing.

Granular material is the same test at a smaller scale, and it is much less forgiving. Salt, ground coffee, powder and sand separate into individual paths the moment they leave a hand. When a model gets it wrong the pinch falls as one soft sheet, more like fabric than like grains, and lands without scattering.

• Volume that is not conserved: the glass fills faster than the stream delivers
• A stream that never breaks into droplets anywhere along its length
• A surface that stays flat while something is entering it
• Grains that land as a single mass, with no bounce and no spread

Look at

the leading edge of the pour, and the surface where it lands

Verdict

Re-roll. There is no post fix for volume.

Cloth, hair, and the frame of release

Fabric weight is legible in exactly one thing: how far a wave travels along the cloth before it dies. Heavy fabric moves in a few large folds and settles fast. Light fabric ripples across its whole area and keeps going after the hand has stopped. A generated garment often drapes correctly in a still and then moves at the wrong weight the moment anything touches it.

Hair breaks the most quietly of anything on this list. Follow one lock across ten frames. Strands should stay a countable group that keeps its own volume; when they merge, split, or grow between frames, the shot has been improvising and you have found the seam.

Gravity is easiest to catch at the exact frame of release. Something leaves a hand and should begin falling in the very next frame, slowly, then faster. A held beat before the fall is the most common single physics defect we see, and people almost always describe it as slow motion when what is actually missing is acceleration.

• A wave that travels the whole length of a garment too heavy to allow it
• Hair that changes strand count between frames
• An object that hangs for two or three frames after the fingers open
• A fall at constant speed rather than an accelerating one

Look at

one lock of hair, one fold of cloth, and the single frame of release

Verdict

Re-roll, unless the shot can be cut before the release.

Hands on objects, contact shadows, occlusion order

Contact is a two-way event. A hand takes a cup and the fingers flatten a little, the cup does not, and the skin changes color where the pressure is. Generated contact is usually one-way: a hand deforming around nothing, or an object held by fingers that never closed.

The contact shadow is the fastest read in the frame. There should be two of them, not one: a tight dark shadow where the object meets the surface, and a wider ambient one further out. The offset between object and shadow scales with the object's own width, which is exactly why one shadow recipe copied onto a bigger object immediately reads as pasted on.

Occlusion order is the cheapest test here and the one almost nobody runs. Take the two things nearest the camera and confirm the nearer one covers the further one for the entire shot. Generated footage swaps them across a cut more often than you would expect, and once you have seen it you cannot stop seeing it.

• Fingers that curve around an object without ever touching it
• One soft blob of shadow under the whole object
• A shadow that keeps the same offset as the object gets larger
• A near object briefly rendered behind a far one

Look at

the exact pixels where two things meet

Verdict

Shadows are often fixable. Grip and occlusion are not.

Mass and inertia

Weight is never in the object. It is in the body carrying it, and in what happens three frames after the motion stops. Look at the wrist, the shoulder and the opposite hip, not at the thing being lifted. If the person would look identical holding nothing, the weight is a costume.

Inertia is the other half. A full bottle set down hard keeps its liquid moving after the glass has stopped. A heavy door swings past its resting point and comes back. Generated motion tends to arrive at rest on precisely the frame the movement ends, which reads as an animation preset rather than an event. We ease every move we build with smoothstep and nothing else, with zero overshoot and zero bounce, for the same reason: a photographed object has no mass to borrow.

• A carried object that does not tilt the person carrying it
• Liquid that stops on the same frame as its container
• A settle with a spring or a bounce on it, which no photographed object has

Look at

the body, and the three frames after the motion stops

Verdict

Re-roll. Weight cannot be added in the edit.

Reflections, specular highlights, packaging text

A reflection has to agree with its object about position, angle and softness, and it has to fade at the rate the surface roughness implies. Specular highlights have a stricter job than that: they must move when the camera moves and stay put when only the object moves, because a highlight is a fact about the light rather than about the object.

Packaging text is where generated footage gives itself away to your customer rather than to you. A model has never read your label, so it writes something label-shaped. Anyone who owns the product can see it in the first second. This is the one item on the list with a clean fix, and the fix is a photograph of the real pack, dropped in as a card and moved like a card. The related trap is borrowed footage carrying somebody else's product name; we once had to cap every window on a stock clip because a competitor's label became legible above a certain magnification, and the shot we wanted was simply lost.

• A highlight that stays in the same place on the object through a camera move
• A reflection that is a copy of the object rather than a view of it
• Label text legible enough to read, and wrong
• Somebody else's brand name readable in footage you borrowed

Look at

the brightest pixel, and any text taller than about twelve pixels

Verdict

Text is fixable with a real insert. Reflections usually are not.