Chapter 1 · What changes when the camera is a model

Yasser Ashour

10 minutes ago

14 min read

Part 1 · Before the camera

On an AI commercial there is no camera to place, no set to light and no take you can watch being made. There is a model that reads what you write, looks at what you attach, and hands back pictures and sounds that nobody shot. This chapter describes what that changes: what the machine does well and badly, what is left of the director's job (almost all of it), and who on the crew touches which tool.

In this chapter

What a model is, and the handful of words this book uses for working with one
What AI does well on a commercial, and where it still fails
How the director's job changes, and what does not change at all
The crew: who runs which tool, and who signs what
What the industry has settled: hybrid work, curation, disclosure

Before you start. Nothing. "How to use this book" explains the labels on an example and the decisions every project makes at the start.

1.1 A camera that reads

A model, in this book, is a program that makes a picture, a clip, a voice or a piece of music from a description. You do not operate it the way you operate a camera. You write a brief for it (the prompt), attach the pictures or recordings it should borrow from (the references), choose a few settings, and it returns a result. Nobody placed a lens or lit a face. The result is a proposal, and it becomes material only when the director accepts it.

Three differences follow, and the rest of the book is built on them.

The picture exists before anyone has seen the room. On a shoot the look is discovered on the day, in front of the lens. Here it is decided in words, and in the stills you approve before any motion is made. Every film starts as a still: first a character, then a frame, then a clip made from that frame. If the frame is wrong, everything built from it is wrong, which is why Part 2 (stills) comes before Part 3 (motion).

The reader has no memory and no judgement of its own. Each job starts from nothing: what you attach and what you write is all the model knows. The video models on the host platform the book uses expose no seed (read 28 Sep 2026; Chapter 33), so you cannot ask for the same take twice. And every model drifts towards the average of what it was trained on unless your words pull it somewhere specific. Left alone, it returns the stock picture; the director's specificity moves it off the average.

The cost moves from crew days to attempts, and the cheapest attempt is a still. On 27 Sep 2026 Higgsfield quoted 2 credits for a 2k still on Nano Banana Pro, 6.25 for five seconds of Kling 3.0 at its standard tier with the sound off, and 35 for five seconds of Seedance 2.5 at 720p. The arithmetic pushes the craft forward: iterate on the still, where a correction is cheap, and spend on motion only after the still is signed. Check every price on the day (Chapter 34).

Figure 1.1 — Sign a still, build from it. Prices are Higgsfield's quotes of 27 Sep 2026 for a 2k Nano Banana Pro still, Kling 3.0 standard (5 s, sound off) and Seedance 2.5 (5 s, 720p); the tool chapters give current prices.

A handful of words mean something particular in this book. Here they are once; later chapters add their own.

Word	In this book
Model	the program that makes the picture, clip, voice or music (Nano Banana Pro, Kling 3.0, Eleven v4 …)
Prompt	the written brief a model reads for one job
Reference	a picture, clip or recording you attach so the model can borrow from it: a face, a room, a product, a voice
Job	one request to a model; each returns a result
Host platform	a service that runs many makers' models under one account; the book's is Higgsfield (Chapter 33)
Route	a way of making something: a model and a method ("a start frame on Kling, its sound set")
Keyframe	a still made to be animated by a video model; not a keyframe on an edit timeline
Sign	approve a still, a take or a cut as final; everything later is built from it and checked against it
Running rough cut	the cut that grows as takes are accepted, where every new take is judged in context
Operator	whoever writes and runs the jobs: the director, an AI artist, or an AI assistant working for the director

1.2 What it does well

Worlds you could not afford to build. A street in another century, a product at a scale no lens reaches, an establishing shot of a place you cannot get to: each can be tried in an afternoon, with a speed and flexibility that a location and a crew cannot match.

Looking before you commit. A still costs little, so you can see a look, a room or a face before paying for motion, and change it. Some of the best uses of AI never reach the screen: a campaign's visual world is found in thousands of generated images and then made conventionally. AI plans; the crew delivers.

Holding a face. Consistency is the failure the industry names most often, and the makers' own answer is to attach the approved image and say what it is for. With a signed still as the authority, a face can be held from shot to shot (Chapter 10).

Speech in a dialect. ElevenLabs advises choosing a voice whose accent matches the target language and region, and says the voice still matters to how a delivery comes out, so the casting comes first. For a finished film the voice is a person who signed or a voice you have cleared, and whether the exact take must ship is decided per project (Chapters 21 and 22).

Many versions of one idea. A cut-down, a vertical version or a different opening is cheap to make. Choosing among them is not (Chapter 30).

The paperwork. An AI assistant is good at the work nobody enjoys: logging every take the same way, frame arithmetic, an edit plan that states every source range and the reason for each cut, notes that name a cause (Chapter 29).

1.3 Where it still fails

The failures are as consistent as the strengths, and most have a craft answer. Read down the right-hand columns and a pattern appears: almost every fix is ordinary film craft, stated precisely.

Where it fails	What you see	What the craft does	Chapters
Physics	shadows that disagree with their source; liquids, reflections and fast motion; the tells cluster around physics, not texture	name each light by what it is and what it does; describe what rests on what; cut before a fast move peaks	6, 7, 14
The same person twice	a face drifts between shots, and further along a chain of edits; OpenAI's guide says to restate what must be preserved when a chained result drifts	sign one still of each character; compare every later frame with it at full size; return to the approved parent instead of editing a copy	9, 10
The uncanny human	a face reads as wrong at once, at the size it will be seen	judge faces at viewing size and at full size; design subjects and actions the models render truthfully	10, 15
Crowds and scale	crowds default to uniformity; the scale of objects to one another breaks	clusters rather than exact counts; test a count before a client shot	10, 14
Words in the picture	misspelled text, invented logos, faint printing on plain objects	two layers: in-world text may be generated and checked letter by letter; logos, titles and legal lines are always set in post	10, 30
Arabic spoken by a video model	no video model's own speech is a delivery route for Egyptian or Saudi Arabic	the voice comes from a voice actor or ElevenLabs; each project decides whether the exact take must ship or a new performance will do	21, 22
Hands and physical states	hands holding small objects are among the hardest cases; what rests on what is easily missed	describe the support and the contact; attach a reference photograph of the pose	9, 10
An action that never happens	a start frame that already shows the action finished leaves the clip nothing to do	freeze the keyframe early in the action, each mover's direction written as positions	9, 14
Repeatability	no seed on the video models; every take is new	keep every original file; record every job the day it runs	8, 33
The generic look	"cinematic" pulls a room towards teal and orange; "8K" and "ultra-realistic" set nothing	name sources, surfaces and colours; no quality words	2, 6, 7
Faces that act	an emotion word invites the face to perform	give the character a task, then visible cues in the body and face	15

How a trained eye finds the tells

Creators list the tells domain by domain: skin that is plastic and pore-free; catchlights missing, or in the wrong place for the light; teeth too white and too even; hair that moves as one solid mass; extra or fused fingers and impossible grips; fabric that does not crease or drape; crowds with the same face twice; reflections that do not match the room; water that refracts wrongly; limbs that stretch in fast motion; shadows that disagree with their light. Practitioners also share a review pass. It is not a measured method, but every step is cheap.

A review pass that costs nothing

Pause at random moments during movement and look for a broken frame.
Zoom into the extremities: ears, fingers, eyes, hairlines.
Step frame by frame through every transition and wherever a hand touches an object.
Watch once with the sound off; visual flaws stand out more.
Slow the clip to quarter speed to check eye movement and small expressions.
Check that reflections match the room, that no crowd figure repeats, and that logos on clothing have not distorted.

Two habits from continuity craft matter more than any single tell: judge a frame first at the size it will be seen and only then at full size, and compare every later face with the signed original, never with a copy of a copy (Chapter 10).

What makes AI work look amateur

Most of what separates amateur AI work from professional work is familiar to any director.

Slideshows. A slow camera and subjects that barely move. The cure is a shot that is staged, with a move and an action that mean something (Chapter 5).
Too much in one shot. One camera move and one action per shot; asking for more creates jitter and floating motion. One dominant camera action per phase and one main subject action with at most two secondaries; a longer clip runs in phases (Chapter 14).
No depth. Amateur shots have a subject and a background; professional ones place things in the foreground, the middle ground and the background (Chapters 5 and 7).
The wrong frame rate. 24 frames a second reads as cinema and 30 as video. The rate comes with the route, so read it from the returned file (Chapter 14).
Light by label. Naming a set-up ("Rembrandt", "three-point") hands the model a word it may read differently from you. Name each source by what it is in the room, where it is, how hard it is and what it does to a surface (Chapter 6).
The grade and the pace. Colour grading is the step that most turns "generated" into "filmed", and cutting on action and to the music is what creates momentum. Make the cut work before any grade or upscale is paid for (Chapters 29 and 30).

The model proposes; the director decides. Every result is a proposal until the director accepts it. A signed frame, a locked take and a locked cut are the authority for everything built after them, and a technical read, however careful, is not a verdict.

1.4 The director's job now

Very little of the director's job disappears. A director who won second prize in a February 2026 competition for AI commercials said the key "wasn't just speed — it was having years of experience in advertising that allowed me to move quickly and make strong creative decisions", and warned that the "biggest mistake people make is focusing too much on visuals and not enough on the emotional arc." The directors whose AI work lands begin from a feeling or a question, and the generation serves it: intention precedes generation.

What changes is where the director's decisions are made and recorded. Seven changes run through this book.

You direct in writing. The brief you would give a camera operator, a gaffer and an actor is now written for a reader that takes every word literally and forgets it after one job (Chapter 2).
You sign earlier, and more often. A frame is signed before it moves; a voice take is locked before a face speaks it. Each signature becomes the authority every later step is checked against (Chapters 9 and 10).
You write the checks before the job runs. "Is the first frame of the clip the keyframe?" "Does the box stop against the laptop, not through it?" A check written after the result arrives tends to describe the result.
You judge in the running rough cut. Takes are accepted or rejected in context, and only the coverage the cut is missing is generated next (Chapter 29).
You decide every change of route and every spend. A different model, a dropped reference or a cheaper mode is a new decision, never a quiet fallback made by whoever runs the job. The price is checked on the day, before the job, and the director approves it (Chapters 8 and 34).
You carry the rights. Real faces, voices, music and client material raise questions that no model answers for you, including what a platform may do with what you upload (Chapter 34).
Your eye and ear pass a shot. "The job completed" or "the face held" is not a pass. Dialect, lip sync, performance and emotional timing are judged by the director; an assistant's observation is a technical read until then (Chapters 15, 22 and 29).

None of this is new to a director. What is new is that there is no set on which these decisions happen naturally, so each one has to be made on purpose and written down.

1.5 The crew, and who uses which tool

AI productions run on very different crews, from one person to a hundred. Whatever the size, the same roles appear.

The director writes or approves every brief, signs every frame, locks every take and the cut, approves every spend, and decides every change of route.
The operator writes and runs the jobs: prompts, settings, references, the price check, the job number, the downloaded original, the record. On a small crew this is an AI artist working for the director, the director, or an AI assistant doing it under the director's instruction. Whoever it is, the operator never signs.
The editor builds the running rough cut and the final cut (Chapter 29).
Sound: a performer whose recorded voice goes into a client final; whoever records Foley, the physical sound of props and hands; and the mixer, who balances and places everything. The edit owns every sound (Chapter 21).
A composer or a licensed music library when a client needs ownership, exclusivity or timing exact to the bar, which generated music cannot promise (Chapter 25).
A motion designer for editable graphics: titles, end cards and animated type are built natively, not generated (Chapter 30).

The table shows who touches each tool. It is not the only way to staff a film, but it shows the principle: the operator's hands are on most tools, and the director's judgement is on every result.

Tool	What it does	Who operates it	Who judges the result
Nano Banana Pro, GPT Image 2.5 (Chapters 11, 12)	stills: characters, frames, keyframes, plates	the operator	the director signs every frame
Kling, Seedance, Wan (Chapters 16 to 18)	motion; a video model can also perform a line when a new performance is acceptable (Chapter 22)	the operator	the director, in the running rough cut
Sync Lipsync 3 (Chapter 24)	redraws a mouth to the approved take on a plate	the operator	the director's eye and ear
ElevenLabs (Chapter 23)	Arabic and English voices	the director or a sound person, on the maker's own account	the director's ear, or a native ear; the take is then locked
Suno (Chapter 26)	music	the director generates, to a brief written in advance	the director, against picture and dialogue in the edit
Sound effects (Chapter 28)	recorded Foley, a licensed library, or generated effects	whoever records or generates them	the edit places each sound; the mix sets its level
Premiere Pro (Chapter 32)	the cut, titles, grade and mix automation	the editor	the director decides the cut and the lock
After Effects, Premiere templates (Chapter 30)	editable motion graphics	a motion designer	the director
Upscalers (Chapter 32)	finishing, only after the cut is locked	the operator or the editor	the director, in motion

If the operator is an AI assistant. Be precise about what it does not do, because an assistant can sound more certain than it is. It logs only frames it has actually sampled and audio it has actually heard. It never judges Egyptian or Saudi delivery, lip sync or acting with authority. When it could not run a step in an editing tool itself, it reports the step as "implemented", never as "executed". A conflict it resolved on its own is returned to the director as a decision. Those limits are not modesty. They are what keeps the director's verdict meaningful.

1.6 What the industry has settled

Three things are settled enough to plan on.

The strongest work is hybrid. Landmark AI commercials are rarely one model from start to finish. They combine generated material with conventional compositing, effects repair, grading and sound built in the edit. That is the shape of this book's route: stray printing and logos come out in post rather than being prompted away, and every title, subtitle and legal line is set in the edit (Chapter 30). AI can also be the idea rather than the shortcut. Orange's "WoMen's Football", by the Paris agency Marcel, won the Film Grand Prix at Cannes Lions in 2024: footage that seems to show the French men's team turns out, through a deepfake face swap, to show the women's, and without the AI there was no idea.

Curation is a job. Every take that is used comes from several that were not. A clip is the most expensive attempt you can make, so the book signs the still first, builds the clip from it and repairs the still rather than re-rolling the clip. Every take is still judged in the running rough cut, and failed runs are kept so that a route's real success rate stays visible (Chapters 29 and 33).

Disclosure is expected, and fabrication is not tolerated. Cannes Lions first required entrants to disclose the use of AI in 2024. In June 2025 it withdrew a Creative Data Grand Prix after AI-generated and manipulated content in the case film was found to have simulated real-world results; the agency's co-president resigned and two more campaigns were withdrawn. The line the industry drew is clear: AI may be in the work, and is never used to fabricate results. Chapter 34 covers the rights questions a director actually meets: real faces, voices, music, references and what a platform keeps.

Four principles come out of this chapter and govern the rest of the book.

Story first, then tools. Part 1 teaches concept, treatment, shot design, light and production design before any model is opened.
Still first, then motion. Sign the frame where a correction is cheap, then pay for motion (Parts 2 and 3).
The edit is where the film is made. Takes are judged in a running rough cut, sound is built there, and finishing waits until the cut reads (Parts 4 to 6).
Record everything, change nothing silently. Every job is recorded the day it runs, every change of route is the director's decision, and every price is checked on the day (Parts 1 and 7).

What to remember

A model reads your brief once, literally, with no memory; everything it needs must be in the prompt or attached to the job.
Left alone, a model returns something close to the average; the director's specificity moves it off the average.
The cheapest attempt is a still: sign the frame before you pay for motion, and repair the frame rather than re-roll the clip.
Most AI failures (physics, faces, words, actions that never happen) have ordinary craft answers, stated precisely.
The model proposes; the director decides. Only the director signs a frame, locks a take or locks a cut.
The operator writes, runs and records every job; the director approves every spend and every change of route.
The strongest work is hybrid: generated material, finished with compositing, grading and sound built in the edit.
AI may be in the work, never used to fabricate results; disclosure is now expected.