Experts are learning the model's language · FinalGrade
FINALGRADE
Perspective

Experts are learning the model's language

Professionals reword prompts until they hit the model's phrasing. The tools should learn the craft's vocabulary, not the other way around.

A cinematographer can tell a camera operator "slow push-in, 35mm, hold focus on her eyes" and get the shot. Every word in that sentence means something precise. "Slow" sets the speed of the move. "Push-in" names the move. "35mm" fixes the perspective and how much of the room wraps around her. "Hold focus on her eyes" tells the focus puller which plane matters for the whole take. Nobody on the crew needs a reference image or a second explanation.

It took a hundred years of film sets to make that language exact. It is exact because money and time depend on it. A misunderstood note costs a take, a lost hour of light, or a reshoot.

What we asked filmmakers

Ask an AI video model for the same shot and something different happens. We wanted to know how often, so before building the benchmark we surveyed 27 working filmmakers who already use generative video tools in their work.

81% said generative models are missing key levers they rely on every day. More than 93% said the models frequently misunderstand their directions. And when we asked what they do about it, most described the same workaround: they reword the prompt into language they think the model understands.

Figure 1Of 27 working filmmakers who use AI video tools
81%

say generative models are missing levers they rely on every day

93%+

say models frequently misunderstand their directions

FinalGrade survey of 27 working filmmakers who already use AI video tools. Each dot is one respondent.

The workaround is the problem

That workaround is the part that matters. It means the experts are adapting to the tool.

They drop "35mm" and try "cinematic." They drop "4:1 ratio" and try "moody." They stop saying "rack focus on the line" and start describing a mood they hope will produce a focus change. They are learning a new vocabulary that was written without them, by people who have never been on a set.

Figure 2On set, a direction takes one step. With today's models, it takes a loop.
ON SET "Slow push-in, 35mm,hold focus on her eyes." CAMERA OPERATOR The shot WITH A MODEL "Slow push-in, 35mm,hold focus on her eyes." Model A missSTATIC, FOCUS WRONG Reword: "cinematic, moody"
Illustration of the workaround respondents described: rewording precise direction into vaguer language until the model responds.

Each rewording loses information. "Cinematic" can mean shallow focus, a wide aspect ratio, a slow move, warm contrast or all of them at once. When a director says "35mm," there is one correct answer. When a director says "cinematic," there are thousands, and the model gets to pick. The professional is no longer directing the shot. They are negotiating with it.

Why this is backwards

The language of filmmaking exists for a reason: it lets a director, a DP and a gaffer agree on a picture before anyone turns on a light. It is a shared specification. A tool that understands it is a tool a professional can direct. A tool that only understands "cinematic" can only be nudged.

This matters for the labs as much as for filmmakers. The people who will pay the most for AI video, and use it most seriously, are the ones who already know exactly what they want. If those users have to unlearn their craft to use the tool, they will use it for mood boards and previs, and not for the shot that ends up in the cut.

What a precise term carries

It is worth spelling out how much a single word of set vocabulary carries, because that is exactly what gets lost in the rewording.

Term on set What it specifies What "cinematic" leaves to chance
35mm Perspective, how much background wraps around the subject, how faces render up close Any lens from wide to long
Slow push-in Direction of travel, speed, that the camera physically moves rather than zooms Whether the camera moves at all
4:1 ratio Two stops between key and fill: a specific amount of shadow Anything from flat to silhouette
Rack focus on the line Which plane is sharp before, which after, and when the pull happens Whether focus changes at all
Motivated from the window Where the key comes from, its color and its angle A flattering light from nowhere

A professional prompt is dense with this kind of information. A model that only responds to mood words throws most of it away.

What the first round says about vocabulary

Our first graded round tested this directly, one term at a time. The pattern is consistent: broad, physical instructions are understood, and precise ones are not. A slow pull back to reveal a room was delivered in 95% of graded clips. Asked as a note on an existing clip, "loop light him," one of the most basic portrait patterns in lighting, was delivered in none. Split light was delivered in 7%. A slow push in that had to land on a specific frame was delivered in none.

These are not obscure requests. Any working gaffer or camera operator would execute them without a second thought. The gap is not in what the models can render. It is in what they understand.

Where the fix starts

The fix starts with measurement. You can't teach a model a language until you can test whether it speaks it, one term at a time.

That is what FinalGrade does. We write prompts in the craft's own terms, the way they would be said on set. Each prompt tests one control and holds everything else as simple as possible. Working professionals judge the results blind, side by side, and write notes on what went wrong. The output is not a single score. It is a map of which words a model understands and which ones it ignores.

Our first round, covering camera and lighting, is already drawing that map. Some directions are delivered most of the time. Others almost never. You can read the results on the FinalGrade-Camera and FinalGrade-Lighting pages.

What labs should take from this

For a lab, the takeaway is practical. Prompt understanding is usually measured with generic descriptions of scenes. Professionals don't write that way. If a model is going to be used by people who make films, its understanding has to be tested in their language, term by term, with the people who use those terms every day judging the result. That is a measurement problem before it is a training problem, and it is the one FinalGrade is built to solve.

Survey: 27 working filmmakers who use AI video tools, conducted by FinalGrade before the first graded round.

Keep reading