AI video analysis

How the analysis reads your video

Plenty of AI can tell you what a video is about. This asks it something much harder: does shot thirty still agree with shot three, down to the ring on her left hand, across the whole clip.

Getting a straight answer to that takes one of the newest video models, the right material in front of it, and a very strict way of asking. Here is every step.

Read the method What it can miss

New runFrom $15 a month · Private to you

Drop a video here, or click to pick one

MP4 · MOV · WEBM · up to 5 GB

01What it writes down

A written record of every shot

Two passes over the whole clip at the same time, each held against the first appearances. Here is what comes out of them.

Every shot: Where each one starts and ends, to a tenth of a second. Everything it finds is pinned to a shot.
Everyone and everything that comes back: People, props and places, named by how they look, with the moment each is first seen clearly.
A review of every shot: Place, time of day, weather, light, background and camera, plus six to ten checks, each marked as set here, still matching, changed or out of view.
Whether it was on purpose: A jacket taken off on camera is part of the story. Only a change nothing on screen explains becomes a problem.

A report's timeline with marks where problems are, and the shot panel underneath listing place, light, weather and every check that held — From a real report. The panel under the timeline lists what was read and checked in the shot you are on.

02The method

Seven steps, one sitting

From the moment you drop the file in to the moment the report is saved.

Film frames laid out on a lightbox with a loupe, one frame circled in orange grease pencil — The old way: every frame on a lightbox, a loupe and a grease pencil. Same job, done across your whole clip in a couple of minutes.

01
Screened before upload
A handful of frames are read in your browser before a single byte is stored. Material this site cannot process is turned away right there, and never uploaded.
02
A copy made for reading
Your original goes into private storage untouched. Straight after, your browser builds a compact copy of the whole clip and pulls up to 48 sharp stills from the original at known times. Video models sample roughly a frame a second at low detail. The stills are how this one reads rings, buttons, hands and small print.
03
Watched in one sitting
The copy and the stills go to one of the newest video models in a single request. It runs in a mode where it can go back over a stretch of the clip and look again before it answers. Shot forty is still being compared with shot three in the same sitting.
04
Two reads, side by side
One read follows the people: faces, hair, every garment, what each hand holds, where they stand. The other ignores the people and follows the room: the furniture, which side of the table the record is on, how many glasses there are, who is left of whom, the shadows, the doors. A read that is busy with faces walks straight past a record that moved across a table, so the room gets a read of its own.
05
Held against the first clear look
Every recurring person, prop and place gets a reference, which is the first time it is seen clearly. Every later appearance is compared with that reference directly, so small drifts cannot stack up from shot to shot.
06
A fixed form
The reader fills in a fixed form with fixed categories and its randomness turned all the way down. Every detail it marks as changed has to come with a problem attached. If one comes back without, it gets turned into a problem in code, so nothing it saw goes missing.
07
Finished in code
The two reads are merged and any problem found twice becomes one. Every timecode, including the ones written inside sentences, is put back on your original's clock. The frame it compared against and the frame that disagrees are cut from your original and saved with the report.

03The reference

First appearance sets the reference

Every later shot gets held up against the first clear look, never against the shot just before it. Small drifts cannot stack up, because there is only ever one frame to agree with.

The model is never asked to keep score, and it never gets to answer in loose prose. That is on purpose, and it is where the research has landed: VideoArgus (arXiv 2608.05485) sets its questions before it looks, FIRM-Video (arXiv 2608.21839) writes the checklist first, and The Low Frequency Trap (arXiv 2608.06361) shows what happens when you let a model do the counting.

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.

04The score

Worked out in code, the same way every time

The score is worked out from the list of problems. It starts at 1.00. A serious problem, like a face that changed or a person or object that appeared from nowhere, costs three. A garment, a held object or something that moved costs two. Light, angle and distance cost one.

All of it is measured against how many shots the clip has, because more shots means more chances to break, so the same problem costs a long cut less. Put the same findings in and you get the same number out. The verdict line is that number put into words.

Worked example12 shots

Start: 1.00
highGlass on the table: 3
mediumHer coat: 2
mediumGold record: 2
lowRain on the window: 1
mediumTable lamp: 2
lowShadow side: 1
Problems, weighted: 11
Room to go wrong, 2 × 12 + 6: 30

1.00 − 11 ÷ 300.63

05The report

What you end up looking at

One screen. The score, your video with a thick timeline marked wherever something breaks, and a panel that follows the playhead: the shot you are in, everything checked there, and any problem with its two frames and its fix.

Shot 05 of the rain scene: the coat is brown, the glass and the record are gone from the table — A report on a re-created scene. Every mark on the rail is a problem, and the panel follows the playhead. High Medium Low

A report on a re-created scene. Every mark on the rail is a problem, and the panel follows the playhead. High Medium Low

The problem list: five problems, each with a timecode, a shot number, one sentence and how obvious it is — The problem list from a real report, with a button to copy every fix at once.

How a fix is written

01The shot and the time
02What is on screen
03What was set earlier, and where
04Whether the story explains it
05What to change, naming the frame to match

Copy one or all of them. Published runs appear on the live benchmark.

06Limits

What it can miss

Your original never goes to the model. The reader gets the compact copy and the stills, so something that is wrong for less than a second can fall between frames, and the smallest details are only as sharp as the nearest still. If your browser cannot build the copy, the run stops and tells you, rather than guessing from part of the clip.

It also does not know what your story intends, so a change you meant can come back as a problem. You overrule it in a second, because the frames are right there. That is the point of showing them.

Reading continuity is one of the hardest things you can ask a video model to do. In TimeCatch (arXiv 2608.23474) the best model noticed two swapped frames 57 percent of the time, where people got it 75 to 92. In Order Matters (arXiv 2608.10908) judges gave a shuffled sequence the same marks as a coherent one. So everything here is built to make the model show its work: the stills, the second read, the fixed form, the frames on every card and a score worked out outside the model.

Two swapped frames, noticed

Best video model57%

People75 to 92%

From TimeCatch (arXiv 2608.23474). The gap is why every card carries its frames.

Benchmark

About the benchmark

The benchmark runs the same briefs through every generator, and each entry goes through this exact check with its prompt published. One caveat worth saying out loud: the reading is done by one company's video model, so the ranking is only as fair as that model is. A second, open judge is on the list.

See the benchmark

In preparation

A paper, eventually

I am putting together a public set of timestamped continuity mistakes in multi-shot generated video: how often this check agrees with a person, how steady it is run to run, and how many of its fixes actually solved the problem once someone made them. It will be linked here when it exists.

Check your clip before anyone else watches it

Sign in with your email, drop the file in, and the report comes back in a few minutes for most clips. Plans start at $15 a month for 15 minutes of video.

Check a video Read the method

How the analysis reads your video

A written record of every shot

Seven steps, one sitting

Screened before upload

A copy made for reading

Watched in one sitting

Two reads, side by side

Held against the first clear look

A fixed form

Finished in code