Guide
Why AI video looks floaty
Weightless, slightly slow, everything gliding. It is the most reliable tell that a clip was generated, and it comes from several things at once — some of which you can prompt away and some of which you cannot.
Cause one
Nothing in the model knows about weight
A video model has no physics. It has never computed a mass, a force or an acceleration. What it has is an enormous number of examples of how things tend to look as they move, and it produces something statistically similar to those.
Weight, in footage, is mostly about acceleration — things speed up and slow down sharply, a foot lands and stops dead, fabric lags the body then catches up. A model producing a plausible average of motion smooths exactly those moments, because sharp changes are the least predictable part and the safest guess is always the middle.
So you get movement with the accelerations sanded off. Everything eases in and eases out. That reads to a viewer as low gravity, even though nothing in the frame is obviously wrong.
Cause two
Smoothness is rewarded, and smoothness is the problem
Frame-to-frame flicker is the most obvious failure a video model can have, so a great deal of effort goes into suppressing it. Anything that keeps consecutive frames consistent gets rewarded.
The side effect is that genuinely abrupt change — which is what an impact, a hard cut of movement, or a fast gesture is — looks a lot like the artefact being suppressed. Pushing down on one pushes down on the other, and the result is motion with no sudden events in it.
Many clips are also generated at a low frame rate and smoothed up to a higher one, which adds another pass of averaging between what was actually generated. The interpolated frames are, by construction, the midpoint of two others — and a sequence of midpoints is the definition of gliding.
Cause three
Missing motion blur
Real footage is made of exposures, not instants. A shutter is open for a fraction of a second and anything moving during that time smears across the frame. We are so used to this that we read its absence as wrong without being able to say why — it is the reason high frame rate film looked strange to audiences, and the reason stop-motion has its particular quality.
Generated frames often have far less of it than a camera would have produced for the same movement. Each frame is individually crisp, the sequence is individually consistent, and the whole thing looks like very good animation rather than footage.
This one is partly fixable after the fact. Adding motion blur in an edit, or shooting the clip with a subject speed that suits the frame rate, closes some of the gap.
What actually helps
Direct around it rather than prompting at it
Ask for less motion. The floaty quality scales with how much is moving. A clip where a person breathes, blinks and shifts their weight is extremely hard to catch out; one where they walk across a room and sit down is obvious. Small motion is not a compromise here, it is the shot that works.
Put the motion in the camera. A slow push or a slight handheld drift over a nearly still subject reads as cinematography and gives the model much less opportunity to invent physics it does not have.
Avoid contact. The giveaway moments are almost always where two things meet — a foot on floor, a hand on a surface, an object set down. Those are the moments that need an acceleration to look right. Frame them out or cut before them.
Cut sooner. The quality is cumulative to a viewer; three seconds rarely trips it and eight usually does. This is the same argument the clip length guide makes from the drift side, arriving at the same advice from a different direction.
Prompt words like “realistic physics” or “weighty” do very little, for the reason the top of this page gives: there is no physics to switch on. They are a description of a style, competing for the same finite attention as everything else you asked for.
Keep reading
The rest of the series
- Why hands and faces come out wrongTwo different failures with two different causes - and why asking for perfect hands does nothing.
- Why the same prompt gives a different pictureA seed is one of five variables. Why fixing it is necessary but not sufficient for reproducing an image.
- How long can an AI video actually be?The per-clip ceiling is real; length comes from assembly, and assembly has its own failure mode.
Where this comes from
Unlimited Studio is built around this one problem
Everything above is what the product does: a character is locked from eight reference angles, each generation conditions on whichever angle its prompt implies, a score checks the result against the reference, and a film is shot one frame-chained shot at a time. What it will and will not generate is set out in the content policy.