A film made in Optra Studio is a few hundred shots. An episode of our animated series is about two hundred. Each one is a few seconds of picture — a girl at a gate, an old man turning his head, a parcel set down on a bench — and each one is made the same way, by the same road, with the same checks at every step.

This post walks that road for one shot, from the line of script it comes from to the moment it plays in the finished cut. It is written for someone who has never seen the engine: what each step makes, what each check asks, what the words on the dashboard mean when a shot says passed, being fixed, stuck or held, and how long a shot ought to be.

The road in one picture

A shot exists in four forms, made in this order. Each form is checked before the next one is made, because a mistake caught in words costs a second to fix, and the same mistake caught in a finished clip costs a render.

1. Words. What the shot shows, who is in it, how they move, and what is said.
2. A picture. A still image drawn from those words: three versions, one chosen.
3. A clip. A few seconds of video made from the chosen picture and the shot’s voice.
4. A place in the cut. The clip trimmed, joined to its neighbours, with voices, background sound and the mix laid under it.

Between the forms are the checks. The rule that runs the whole studio is simple: if a shot passes its checks, the engine goes ahead and makes the next form by itself. Nobody approves shots one by one. When a finished film falls short, that is treated as a hole in the checks, and the fix goes into the checks — never a person put back in the loop to catch it by eye.

Step 1: the script

Everything starts with the script: the story, what each scene is for, the dialogue, the narration. The script is written by a frontier language model with a human director, because this is the one job where taste decides everything and where we measured the difference: judged by a panel on the same scale, scripts written this way scored around 6 to 7 out of 10, and scripts from the self-hosted model we use for bulk work scored 2 to 3, and prompt engineering moved that by a fraction of a point.

The script is a few thousand words for an episode. It says nothing yet about cameras or pictures.

Step 2: the board — turning the script into shots

The board breaks the script into shots. This is bulk work in a fixed format against rules a program can check, so it is done by our own self-hosted model, Qwen, running on our own GPUs. Each shot on the board carries:

The picture	What the frame shows, in plain words: the place, the light, who is in it, where each person is, which way they face, what they wear and hold. Characters are written as named tokens, so the engine can put each character’s fixed look into the words it sends to the image model.
The motion	What moves during the shot: she lowers the parcel onto the bench, he turns toward the gate.
The cast	Who is in frame.
The line	Who speaks and what they say, if anyone does — or a narration line laid over the picture.
The length	How many seconds the shot is planned for. As step 5 shows, the voice has the last word on this.

Before any of this is drawn, each character has a look written in words — age, build, hair, clothes, the colours that are theirs — and a character sheet: a reference picture drawn from those words, three versions, the best one chosen by a vision check. Every still of that character is drawn with their sheet as a reference, so the same person looks like the same person from shot to shot. Writing a look out in words turned out to matter as much as the reference picture: on one short film, a character’s likeness failed on every take of a shot until his look was stated plainly in words, and then passed on all three.

Step 3: checking the words before anything is drawn

This is the cheapest place to catch a mistake, so every shot’s words are reviewed before a single pixel exists. Qwen reads each shot and answers four questions about it:

Check	The question it asks	What it catches
Starts right	Does the picture show the moment the movement begins?	A picture that already shows the end: a woman already seated in a shot whose motion sits her down. The video model would have nothing to animate.
Can be seen	Is anyone being looked at somewhere the looker can see them?	A man facing the camera while the person he looks at stands behind him in the background.
Makes sense	Is anything in two states at once?	A cap worn on the head and held in the hands in the same frame.
Looks at the listener	Does a speaker look at the person they are talking to?	Someone speaking a line to a friend while written as staring straight ahead.

Three rules make these answers trustworthy, and each one was added after the check got something wrong:

• A fail must quote the shot’s own words, exactly. If the model says a shot fails, it has to copy the words that prove it. If the quote is not in the shot, or says nothing about the question asked, the fail is thrown out — and the dashboard says so, showing the fail it gave, rather than passing it silently. Before this rule, the check invented faults.
• For the two hardest questions, the model reports facts and code decides. For starts right, the model writes where the movement starts, where it ends, and which of the two the picture shows; the program fails the shot only if the picture shows the end. For can be seen, it writes which way the looker faces and where the other person is; the program applies the rule. Early versions let the model decide, and it failed shots with reasons that described a pass.
• A question that does not apply is not asked. A shot where nobody speaks has no listener to look at, and the dashboard shows that as not asked, not as a pass.

Beside these four, a set of plain rules runs over the words with no model at all: that a character wears only their own colours, that the time of day does not jump inside a scene, that two shots do not ask for the same picture word for word, that a narrator is not heard over another character’s face looking at the viewer, and more. Some of these block a shot; others are reported until they have been measured.

What passed, failed, stuck and held mean

Each shot carries one status for this stage. On the dashboard, every shot also lists each question it was asked and what the answer rested on — for a pass as well as for a fail, because a check that only shows its failures cannot itself be checked.

Status	What it means	What happens next
Passed	Every question that applies passed.	The shot moves on to be drawn, by itself.
Being fixed	A check failed.	Qwen rewrites that one shot’s words, given the reasons it failed, and the checks run again — up to three rewrites.
Stuck — needs a note	It still failed after three rewrites.	The engine stops rather than loop forever, and waits for a person to write one note on what to change. The next rewrite carries that note.
Checker unsure — needs a look	The checker’s answers do not settle the question.	The engine stops itself and a person looks.
Failed — held, not rewritten	It failed, but the whole project is on hold by its settings.	The checks run and report, and nothing is rewritten or drawn until the hold is lifted.
Not checked yet	The words changed since the last check, or it has never been checked.	It is checked on the next pass.
Not checked — the checker did not answer	The model call failed.	It is asked again. No answer is never counted as a pass.

When every shot of a board has passed, the engine records its own approval of exactly that version of the board — every shot’s words, fingerprinted — and starts drawing. If anyone changes a shot afterwards, the fingerprint changes and that approval no longer covers it.

Step 4: the picture

Each shot is drawn three times, with the same words and three different random seeds, with the character sheets of everyone in frame as references. Then a vision model — a small one that looks at pictures, running on its own GPU — screens every take and the best one is picked.

The screen asks every picture a numbered list of questions, one line each:

1. How many people are in this picture, counting anyone partly out of frame or far in the background? (Checked against how many the shot asks for.)
2. Are there any letters, words or numbers anywhere? (Stray lettering, signs and flags nobody asked for are the most common thing an image model adds.)
3. Is anyone missing an arm, a hand or a leg, or has one too many?
4. Is there a stray tooth, a blob or a mark on a face?
5. Does the same person appear twice?
6. Is any object drawn wrongly — broken, melted, doubled, with a part missing or extra, like a pot with two spouts?
7. …and then the shot’s own facts: up to six short statements written from the shot’s words by Qwen (a woman in a yellow raincoat, she holds a parcel, it is night), each asked as a yes-or-no question.

The list is numbered for a reason we learned the hard way. When the questions were written as labelled lines followed by numbered facts, the small vision model answered only the numbered lines, three times out of three, and the program reading its answers threw every reply away: 183 of one episode’s 194 stills read as unanswered. Asked as one numbered list, the same model answered 72 of 72 test questions correctly, identically across three runs, and the larger model we fall back to read the same list too. And a still that gets no readable answer now blocks its shot, instead of being counted as fine.

Two more rules protect the screen from itself. A fact the model answers wrongly is asked again, alone, in a fresh call; if it still disagrees with the words, it is asked differently: the model describes the picture in its own words, and Qwen judges from the description, quoting it. And the screen is measured, round after round, against a person’s verdict on pictures they have actually looked at. On our series it went from about a quarter of its flags being right to 64% in three rounds of fixes, and it reports rather than blocks wherever it has not yet earned that trust.

If the picture itself is wrong but the words are right — letters on a wall, an extra hand — the shot is redrawn: same words, new seeds, three new takes. If the words are wrong, the shot is rewritten. They are different remedies for different faults, and the dashboard offers both.

Step 5: how long a shot should be

The board plans a length for every shot, but for a shot where someone speaks, the voice decides. Every line is voiced first, in every language the film is made in, and the voice file is measured. Then:

A speaking shot	0.3 seconds before the line starts, then the line itself, as long as the recording actually is, then 0.5 seconds after it ends — rounded to whole frames, and never shorter than 1.5 seconds. The clip is never shorter than what is said in it; in the final cut the shot is trimmed back to exactly its line.
A wordless shot	Keeps its planned length: a silent beat is a storytelling choice. But a wordless shot longer than 3 seconds with no narration over it is flagged as dead air.
The ceiling	Never longer than the video model has been shown to render cleanly. On an RTX 5090 (32 GB), the video model breaks down past about 265 frames — roughly 11 seconds — so longer clips go to a card with more memory.

Why measure instead of plan? Because planning got it wrong at scale. Our series’ old boards gave every line a formula — a little over half a second per word plus two and a half seconds — while the voices said six words in about three seconds. Of 2,290 speaking shots, 2,255 were longer than their line needed: about 46 minutes, across the series, of characters standing still after they had finished speaking. Measuring the voice and trimming to it removed that.

In short: a speaking shot is its line plus 0.8 seconds, and at least 1.5; a wordless beat is two or three seconds, longer only with narration over it; nothing runs past what the model renders cleanly.

Step 6: the voice

Every line is voiced in each character’s own voice, in each language. Several takes are made and the best clean one is kept; a take with a click in it is voiced again, because one click can cap the loudness of a whole episode. Then a speech-recognition model listens to every take that will be used and checks that it says the words it was written to say — a take that says something else, or an old draft of the line, is caught, and the cut waits for it.

A line’s voice file is fingerprinted by its words, its voice and its language. Change a word of the script and it becomes a different voice job, voiced fresh; an old take can never be laid under a new line by mistake.

Step 7: the clip

The clip is made from the chosen picture. What drives it depends on the shot:

• A line spoken on screen is lip-synced: the voice itself drives the video model, so the mouth moves to those words. That means one clip per language — the mouth has to say that language’s words.
• Narration, a line spoken off screen, or a wordless shot is animated without a voice, and the same clip is shared by every language. The narration is laid over it later, in the cut.

A clip’s name is a fingerprint of everything it was made from — the picture, the voice, the motion words, the length, the size, the model and the seed. Change any of them and it is a different clip. That is how the engine can never cut a film from a clip left over from an older version of a shot.

This is the long step. On four generation GPUs — three RTX 5090s (32 GB each) and an RTX PRO 6000 (96 GB) — the stills for an episode of about two hundred shots take around an hour, and the clips take most of a night.

Step 8: reading every frame

A finished clip is broken into its frames — 24 per second — and every frame is shown to the vision model. Not a sample: every frame. Sampling is exactly how defects used to reach finished episodes: a black bar across the first second and a quarter of a shot shipped twice past a check that sampled twelve frames of it. At 24 frames a second, a defect has nowhere to hide.

Each frame is asked, in one call, whether a hand, a finger, a foot or a face is malformed, and how many people are in it. Then the answers are read as runs, not counts:

• One odd frame in a whole episode is the model blinking.
• Six in a row is a quarter of a second a viewer can see: reported.
• Twelve in a row is half a second: the clip fails and is made again.

The vision model runs on its own card, an RTX A5000 (24 GB), which does nothing else. With twelve questions in flight — measured as the point where the card is saturated, and more only adds waiting — it reads about 6.3 frames a second, a little over an hour for a whole episode, while the other cards keep rendering.

Step 9: the cut, and the check of what ships

The cut joins the clips in order, each speaking shot trimmed to its line. The voices are laid on one track, sample-exact, and the picture and sound are joined once — joining shots that each carry their own sound makes the voice drift against the picture over a long film. Each scene gets its own background sound, a room or a street or rain, dipped under every line. The mix is set by the loudness of the speech, not of the whole track, so a loud background cannot make the voices quiet.

Then the whole film is checked again, as a finished thing. Every part of this final check reads the film as it ships:

• Every still, screened again with its shot’s facts.
• Every frame of the finished cut file read at 24 frames a second. Not the clips it was cut from: the cut. A trim, a join, a held first picture or an end card exists only there. A frame whose picture is byte-for-byte one already answered is not asked twice — the other language’s copy of a wordless shot, or a re-cut of unchanged clips — and the dashboard says exactly how many of the cut’s frames were read, of how many, and how many were asked this time.
• Every voice, heard saying its words.
• Every shot that should make a sound, checked to be audible in its own place in the cut — per shot, so one silent shot cannot hide in a noisy scene.
• Long silences, picture and sound ending together, the loudness.
• A panel of story editors reading the cut in order and scoring it, reported until that score has been measured against what people think.

Every finding is listed with its shot, its language, the second it happens in the cut, and a button that plays the cut from exactly there.

When a finished film falls short

Sometimes a person watches a film the checks passed and sees something wrong: a parcel floating in mid-air, a smeared hand, a speaker looking into the lens. That is not a reason to put a person back in the loop. It is a hole in the checks, and it is closed as structure, in three questions asked every time:

1. Would a smaller version still pass? One shot instead of eleven; one frame instead of a second.
2. Does the fix hold for films not made yet? Every property a finished film must have is on one list, and every kind of production must answer for each one: checked, exempt with a reason, or missing in writing.
3. Is the thing checked the thing that ships? Not the step, not a file on disk — the film a viewer watches.

To close a hole, a person marks the picture on the dashboard — the exact frame of the cut, or the exact still — as wrong or as right, with the check that should have judged it. The engine keeps that picture, byte for byte the one the check was given, as a test case, and every change to a check is measured against the kept cases before it is allowed to change a verdict: it must flag the wrong ones and pass the right ones, three runs out of three.

What is still being built

Three properties are on the list and not yet checked, and we say so rather than let their absence read as a pass: that a shot is no longer than its line needs and than the model renders cleanly; that nothing is mixed over speech; and that a scene keeps moving. Two more are being built from what people have seen in finished films: that the motion a shot is written with actually happens between its first and last frame, and that clothes, props, posture and place stay continuous from one shot to the next.

The whole road, once more

Step	What is made	What checks it
Script	The story and every line	A panel of story editors
Board	Each shot in words	Four questions per shot, plus plain rules; rewritten up to three times
Picture	Three takes per shot, one chosen	A numbered screen of every take: people, lettering, limbs, marks, doubles, objects, the shot’s own facts
Voice	Each line, each language	Heard back by speech recognition; clicks caught
Clip	A few seconds of video	Every frame read at 24 a second; runs, not counts
Cut	The film	Every frame of the finished file, every voice, every sound in its place, the loudness, the story

Every step makes one thing, every check asks about exactly that thing, and a shot moves on only when its checks pass. When one of them misses something, the miss becomes a new check. That is how a shot is made.