Why Every Film Set Has Someone Watching the Coffee Cup

An AI video ad is assembled from separate generations, and the model that renders shot six has no memory of shot five. If a scarf, a bottle label or the coffee in a cup changes between them, nothing in the model notices. So the first question for any AI video project is an old one: who is watching continuity?

Film crews have lived with that problem for more than a century. Scenes are shot out of order, takes are hours apart, and the audience still expects a glass to stay half full across a cut. Professional sets carry one person whose job is to notice: the script supervisor, long called the script girl.

A job invented in 1918

Writing in Script magazine, Rosanne Welch credits screenwriter Sarah Y. Mason with essentially inventing the craft in 1918, when she worked as the continuity person on Arizona, a Douglas Fairbanks production. NPR’s account ties the need to the moment D.W. Griffith moved the camera in closer and multiple cameras came into use. Once a scene was covered from several angles, every angle had to agree.

The work was often handed to someone from the secretarial pool, and the name “script girl” stuck. Darren Paul Fisher, head of directing at Bond University, writes in The Conversation that early Hollywood typically filled the role with women. NPR reported in 2008 that the job was still done primarily by women.

Generated video has no such person built in. Each generation starts fresh from its prompt and whatever images it is given, with no idea what the previous shot looked like. On an AI project, the script supervisor’s job falls to whoever directs the work, and it has to be assigned on purpose.

What gets written down

An IATSE job description lists the duties: tracking the continuity of hair, makeup, costumes, props and sets, back-matching props scene by scene, taking continuity photos, and helping the director and cinematographer with eyelines and screen direction. ScreenSkills adds the timing of every take and marked-up script pages, delivered daily, so the editor knows what was shot and what will cut together. The tools changed, Fisher notes, from notes and sketches to Polaroids and now smartphones.

Eyelines and screen direction are the invisible part of the job. A character who leaves frame on the right has to enter the next shot from the left, or the walk reads as a turn back. In two separate close-ups of a conversation, one actor looks just left of the lens and the other just right, or they seem to stare past each other.

Food is the classic trap. “People keep eating the food,” script supervisor Steve Gehrke told NPR, and someone has to know whether the plate held 23 peas or 14. In his experience, most continuity errors come when shooting resumes after lunch. The same report lists slips that reached the screen: Julia Roberts’ croissant turning into a pancake in Pretty Woman, Sean Connery’s collar flapping open and shut in The Untouchables, and Dorothy’s dress spotless right after a fall into a muddy pigpen in The Wizard of Oz.

That list doubles as a map of where generated video breaks. Between two generations, a jacket gains a pocket, a ring jumps hands, or an actor walking left to right enters the next shot from the wrong side. Inside one generation, a label can warp as the bottle turns. The script supervisor’s categories make a ready-made checklist for reviewing AI clips shot by shot.

When nobody catches it

In May 2019, viewers of the fourth episode of Game of Thrones’ final season spotted a takeout coffee cup on a feast table, among goblets and drinking horns. Many took it for a Starbucks cup. HBO answered with a joke: “Daenerys had ordered an herbal tea.” CBS News reported that HBO then deleted the cup from the episode, and that two weeks later fans found plastic water bottles near Samwell Tarly’s feet in the series finale.

Fisher lists older classics: a stormtrooper in Star Wars bumping his head on a Death Star doorway, and bullet holes already in the wall in Pulp Fiction before anyone fires. He also describes a candle that seems to “trombone,” shorter then longer from one take to the next.

Many slips are choices. Directors and editors, Fisher explains, are led by the quality of a performance: an actor may give the right look at the end of a take when it belongs at the start of the scene, so the shot moves and the background stops matching. With the budget, a clock can be changed digitally or a coffee cup made to disappear. Often the team takes the hit, and most errors go unnoticed on first viewing.

Generated video slips differently. On set, the usual error is a stray object; in a generated shot, the object itself mutates. Il Sole 24 Ore reported that Coca-Cola’s 2025 AI-made holiday ad drew criticism for trucks that changed shape along the snowy road, and Business Insider noted that their wheels appeared to change in number as the ad went on.

A continuity book for a machine

Most video models build a clip from random noise, steered by a prompt and any images they receive. Nothing carries over from the previous clip unless it is fed back in, so two prompts asking for “a woman in a red coat” will usually produce two different women.

A May 2026 benchmark, EntityBench, by Ruozhen He and colleagues, tracked characters, objects and locations across 2,491 shots drawn from real narrative media. In existing generation methods, consistency “degrades sharply with recurrence distance”: the longer a character or prop is off screen, the more likely it returns changed. The best character results came from storing verified reference images for each one before generation began.

The continuity book, then and now

Script supervisor duties on set and their counterparts in generated video

Script supervisor duties and their counterparts in generated video

On set	In generated video
Continuity photos of every setup	Locked reference images reused in every shot
Wardrobe and prop notes, scene by scene	Character and product reference sheets
Matching action across the cut	Last frame of one clip used as the first frame of the next
Eyelines and screen direction	Gaze and direction of travel set for each shot
Editor’s log and lined script	Log of every generation and the approved version

That is the script supervisor’s method, rebuilt for generation. Each recurring character, wardrobe look, product and location gets a reference sheet, approved once and fed into every shot that needs it. Many models accept a first frame, and some a last, so the final frame of one clip can open the next and keep a hand or a glass in place; the chain carries flaws forward too, so each link is checked. A product that must be exact is often safer filmed and composited. A log records each clip’s references, prompt, approved version and notes such as ring on the left hand, and clips are compared side by side at every cut before upscaling, while a fix still costs one generation.

What to ask of any AI video project

Ask who owns continuity, by name. Ask to see reference sheets for every recurring person, product and place before generation starts, and confirm every shot uses them. Ask for the continuity log, and review stills from both sides of each cut: wardrobe, hands, props, eyelines and direction of travel. Check your product in every frame where it appears, down to the logo, the label and the number of parts. A coffee cup got past the crew of a show watched by millions; in AI video, the cup can also change shape between shots.