AI video continuity

Keeping a generated cut consistent

On a film set someone sits with a notebook so the jacket is on the same shoulder tomorrow. Generated video has no notebook and no set.

Video models make 15 to 30 seconds at a time, so every piece of a long cut is made on its own and the jacket, the room and the weather are up for renegotiation each time. This is how long AI video really gets made, what goes wrong, the working order that keeps a long cut from falling apart, and how to find and fix the shots that broke anyway.

01How it gets made

A long AI cut is a pile of short ones

Today's video models top out at around 15 to 30 seconds per generation. Nobody is generating a five-minute film in one go. What people actually do is make it in pieces, 15 or 30 seconds at a time, and edit the pieces together. Five minutes is ten to twenty generations.

That works right up to the seams. Each generation is made fresh from the prompt and whatever reference you attach, with no idea what the last one looked like. So between generations the jacket changes shade, the face settles a little differently, the window moves walls and the afternoon turns into evening. Every piece looks fine on its own. Played in a row, the seams are where AI video gives itself away.

That gap between generations is the problem VideoConsistency is built to solve, as simply as I could make it. Upload the cut, our agent checks every generation against the first, and the shots that slipped get fixed.

One three-minute AI video

8 generations · 7 cuts between them

A three-minute video made of 8 separate generations, each 15 to 30 seconds long. Something drifted at 5 of the 7 cuts between them.

  1. 10:20The coat comes back a shade lighter
  2. 21:00The record is on the shelf now
  3. 31:30Afternoon turns into evening
  4. 42:05Her face settles a little differently
  5. 52:30The window has moved to the other wall
Each block is a separate generation of 15 to 30 seconds. Every one is made from scratch, so the cuts between them are where things drift. The drifts shown here are an illustration.

02What breaks

What breaks, in order of how often

None of these are exotic. They are the same mistakes a script supervisor writes down on a live set, just arriving faster and in bulk.
  • A glass of red wine and a cigarette in an ashtray on a cafe tableProps
  • An actor fastening his shirt cuff under a tweed jacketWardrobe
  • A street half in low sun and half under cloudLight
  • A car driving left to right on a wet roadDirection
01
The person
A face settles slightly differently every time it is generated. Hair length, a beard, the shape of a jaw.
02
The clothes
A jacket comes off between two shots that are meant to be the same minute. A shirt changes shade.
03
The things they hold
A glass moves hands. A record is on the table in one shot and on the shelf in the next.
04
The room
Furniture shifts, a window appears on the other wall, the light turns from afternoon into evening.
05
Where they stand
Two people swap sides of the frame, so a conversation suddenly turns the wrong way.
06
The weather and the clock
Rain in one shot, dry ground in the next. Long shadows, then no shadows at all.

There is a longer walk through the categories on continuity errors, and the face and wardrobe side has its own page on character consistency.

03The habit

Always compare to the first shot

This is the one habit that changes the most. Pick the shot where a person or a place is clearest, treat it as the truth, and measure everything later against it.

Comparing each shot to the one before hides drift, because every step looks close enough to the step behind it.

The first clear look at the woman at the table: green coat, glass of water, gold record on the table, rain on the windowReference

Shot 02 · 0:08 · first clear look

  • Shot 03 · 0:13Matches

  • Shot 04 · 0:21Matches

  • Shot 05 · 0:314 breaks

  • Shot 07 · 0:44Matches

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

04The workflow

A working order

A filmmaker pointing at one panel on a wall of storyboard sketches while a colleague holds the shot list
Decide what everything looks like before you generate anything. It is a lot cheaper on a wall than in a re-render.
  1. 01

    Pick your reference shot

    Whichever shot shows the person or the room most clearly. Everything later gets held up against it, not against the shot before it.

  2. 02

    Generate against the reference, not the last frame

    Chaining each shot off the previous one is how drift compounds. Small changes stack until shot twelve has nothing to do with shot one.

  3. 03

    Cut first, check second

    Continuity only matters in the order the audience sees it. Assemble the sequence, then look for breaks across the cut.

  4. 04

    Decide what is deliberate

    If she takes the coat off on camera, it is not an error. Keep a short list of changes the story explains, and ignore flags that match it.

  5. 05

    Fix the shot that broke and keep the rest

    Redo only that shot, against the reference frame, then check again. Redoing the whole sequence just gives you a fresh set of problems. Fix this shot in the report does exactly this for you.

05Find and fix

Where the check and the fix save you time

Doing this by eye means scrubbing a five minute cut over and over, holding shot one in your head while you watch shot twenty. That is the part worth handing off.

Our agent watches the whole cut in one sitting, follows the people and the room separately, and holds every appearance against its reference shot, generation after generation. You get a score, a timeline with a mark wherever something stops matching, the two frames that disagree, and a written fix for each mark. Every shot also gets written up with what was checked in it, so a clean result comes with its working.

Then comes the fix. Under every written fix is a Fix this shot button. One of the best frontier video editing models regenerates that one shot against the reference frame, you see a low-res draft first, and the fixed shot drops back into your cut. How fixing works.

If you want the method itself, how our agent watches your video and how the number is worked out, that is on AI video analysis. Published runs sit on the benchmark.

Report · shot review
A report's timeline with marks where problems are, and the shot panel underneath listing place, light, weather and every check that held
A real report. Every mark is a break, and the panel underneath shows what was checked in that shot.

06Your part

What stays your call

Our agent does not know your script. If the coat comes off on camera, it sees that and leaves it alone. If the coat is gone between two shots because a day has passed in your story, it will flag it, and you are the one who knows the day passed. It watches a compact copy of your video with sharp stills from the original, so something wrong for less than a second can fall between frames.

Which shots get fixed is your call too. Nothing is regenerated until you press the button, and you see a draft before you take the full-quality file.

Run it again after every round of fixes. A regenerated shot can fix one thing and quietly break another, and the second run catches it.

Find the breaks and fix them before anyone else watches

Sign in with your email and drop the file in. The report comes back in a few minutes for most videos, with a fix for every break. Plans start at $19 a month for 15 minutes of video and 10 seconds of shot fixes.