Continuity errors in film, and why AI video makes so many more of them

A hundred years of film built a whole job around remembering where the glass was. Generated video threw that away and gets the same errors far more often. Here is what changed, and how to find and fix them.

9 September 20267 min read

There is a job on a film set that most people watching a film have never heard of. The script supervisor sits near the camera with a notebook and a stills camera, and their entire day is spent writing down where things were. How much wine was in the glass. Which hand the cigarette was in. Whether the third button was done up. Which way she was looking when she said the line.

It sounds like the least glamorous job in cinema and it sort of is. It is also the reason films hold together at all.

A continuity binder open on set, handwritten notes on the script page and polaroids of every take on the facing page, a stopwatch and a radio beside it — The notebook. Every actor and prop photographed, every take written down.

Nothing is filmed in the order you watch it

Here is the thing that makes continuity hard, and it has nothing to do with anybody being careless. A scene you watch as thirty unbroken seconds was not shot as thirty unbroken seconds. It was shot as a wide, then a two-shot, then her close-up, then his close-up, then an insert of the glass on the table. The wide might be Tuesday morning. Her close-up might be Thursday afternoon, because that is when the location was free.

So between two shots that the audience reads as consecutive, an entire room has to be put back. The glass at the same level. The chair at the same angle. Her hair parted the same way. The sun in roughly the same place in the sky, which is the one thing nobody can actually control.

Most of the time it works, because somebody wrote it down and photographed it. And still things get through. There are websites with thousands of them. The famous ones are famous because they are in good films made by careful people with a script supervisor on set. That is how hard this is.

Two frames of the same breakfast table. A croissant in the first, pancakes in the second, and the coffee cup has moved. — Two shots of one breakfast, filmed two days apart. Somebody had to remember it was a croissant.

The usual suspects

After you look at enough of these, they fall into a short list. Props, and mostly food and drink, because an actor has to eat and drink through fifteen takes. Wardrobe, where a collar changes colour or a watch swaps wrists. Hair and makeup, because hours pass between the wide and the close-up and hair does not sit still. Screen direction, where a car going left is suddenly going right. Light and weather, when the sun moved or the second unit came back a week later. Where people and furniture are standing. Eyelines. And on long shoots, the season, so a single afternoon starts in summer leaves and ends in bare branches.

I wrote each of those up properly with what it looks like on screen on the continuity errors page. The list matters here for one reason: it is exactly the same list for generated video.

What changes when the shot is generated

Same categories, different cause, and a much higher rate.

The same woman in two generated close-ups. In the second the earring is gone and the scarf is a different scarf. — Generated, five shots apart. Nobody moved anything. The earring and the scarf just came back different.

On a set, the room is real. It persists whether or not anyone is looking at it. Continuity errors happen because a real room got put back slightly wrong. With generated shots, there is no room. Each shot is made from nothing, so there is nothing to put back and nothing to put it back from. The model is not misremembering the jacket. It never had the jacket. It has a description of a jacket and it makes a fresh one every time you ask.

That produces a failure that feels different to watch. A film continuity error is usually one specific thing in the wrong state. A generated one is often a slight difference across everything at once, which is much harder to point at and much easier to miss. Every shot looks completely fine on its own. You only see it when you put two of them next to each other, and by then you have watched your own footage forty times and gone blind to it.

Part of it is simply how long AI video gets made. Video models top out at somewhere around 15 to 30 seconds per generation. So anything longer, which is most things worth making, is a stack of short generations edited together. A three-minute film is six to twelve separate goes at the same people in the same room, and every cut between two of them is a fresh chance for the jacket, the face or the light to come back slightly different. That is usually the moment a viewer clocks it as AI.

The other thing that goes wrong is the way people work. You generate shot two from shot one, shot three from shot two, and so on down the chain. Each step drifts a tiny bit and each step drifts from an already drifted starting point. Twenty shots later the person at the end is visibly not the person at the start, and no single cut is where it went wrong.

Compare with the first shot, not the last one

That drift has a straightforward fix and it is the same trick the script supervisor uses. They do not compare this shot with the previous shot. They compare it with the photograph they took the first time, which is why the notebook exists.

So pick the first clear look at each person, each room and each prop. Call that the reference. Hold every later shot up against the reference and never against its neighbour. Errors stop compounding, because there is nothing to compound against.

The first clear look at the woman at the table: green coat, glass of water, gold record on the table, rain on the window — Every later shot is compared with the reference, never with the shot before it. Shot 05 is the one that disagrees.

Every later shot is compared with the reference, never with the shot before it. Shot 05 is the one that disagrees.

There is one wrinkle. Sometimes something changes on purpose. She takes the jacket off, so of course the jacket is gone in the next shot, and calling that an error is worse than useless. Anything checking continuity has to hold the possibility that the story explains the change, which means looking at what happened between the two shots rather than only at the two frames.

Why I built the check, and then the fix

I got tired of finding these myself, badly, at midnight, after the thing was supposedly finished. Watching your own footage frame by frame is a terrible use of a person and you are the worst possible reviewer of it anyway, because you already know what you meant it to look like.

So I built something to do the notebook part, in the easiest way I could think of. You upload the finished cut and our agent watches your whole video in one sitting, with sharp stills from the original alongside so it has clear frames to compare. Two reads run at once, one for the people and what they wear and carry, one for the room and where everything stands. It works out the shots, picks a reference for every person, prop and place, holds every generation to it, and writes up every shot, including everything that held.

What comes back is a score worked out in code from what it found, a mark on the timeline at every break, the two frames that disagree with their timecodes so you can check the call yourself, and a written note per problem you can paste into your next generation.

Then I made it fix them. Under every note is a Fix this shot button. One of the best frontier video editing models regenerates just that shot against the frame where things were right. You get a low-res draft first, and an automatic check says whether the problem is gone and whether anything else moved. In my own test video a wall that had turned brown came back blue on the first draft. A figure that had swapped sides of the room did not, which is exactly why you see the draft before you take the real file. How fixing works.

The frames matter more than the score. An agent saying there is a problem is an opinion. An agent showing you the two frames is evidence, and you can overrule it in about a second if it is wrong.

Nobody has properly solved this for generated video yet, which is what makes it interesting to work on. Throw your video at it, see what it catches, and press the button on the worst one. If it misses something obvious I would genuinely like to know, because that is how the check gets better.