The One
Small Change

At some point during the production of a television commercial for Gigaclear, a rural broadband company in the United Kingdom, the image archive passed four thousand files.

There was no single moment when the archive became difficult to manage. The files accumulated: roads, cottages, fields, hangars, crowds, skies, rooms, lighting tests, rejected characters and near-matches that needed another pass.

By the end, there were 4,063 images.

That number is useful because it describes the practical reality of AI production. The system does not turn an idea directly into a finished image. It produces many possible images, and someone still has to identify the version that works for the brief.

The Gigaclear commercial was an ambitious AI production by conventional standards. Approximately ninety-five per cent of its visual world was generated or constructed using artificial intelligence. It contained rural landscapes, architectural environments, a large crowd and an orange-suited hero called the Giganaut.

Described quickly, the project can sound as if traditional production has been removed from the process. A national commercial was created largely without travelling to the places it depicted. Locations were generated. Crowds were assembled synthetically. Environments that would once have required scouting, permits, transport, weather planning, catering and a large location unit were made inside computers.

But the project still took roughly a month.

It still required a director, an agency, actors, artists, reviews, approvals, Cinema 4D, Redshift, After Effects, compositing, rotoscoping, paint work, grading, cleanup, and repeated conversations about changes that appeared minor until someone attempted to make them.

The commercial did not demonstrate that production had disappeared. It showed that generative work had become another department within it, with its own requirements for selection, control and finishing.

01 - The Demonstration and the JobThe First Image Is Almost Never the Job

Public discussion of generative AI often concentrates on what the technology can make. That is understandable. A sentence enters a text field and, shortly afterward, a desert palace, an alien city or an expensive-looking perfume advertisement appears. In a production environment, however, the first image is only the beginning of the work.

The first image is often remarkable. It is also almost never the job.

For Gigaclear, that distinction showed up in the first review. The question was no longer whether the model could produce a rural landscape or a crowd. It was whether the chosen version could survive the next request without changing everything around it.

These sound like adjacent requirements. In practice, they are almost opposites.

Generative models are designed to invent. Production is organized around decisions. A model prefers to reinterpret an instruction each time it receives one; a production needs it to remember that the previous interpretation has already been approved. The model wants to continue exploring. The client would like the exploration to stop, except in one carefully selected area.

This conflict often appears in a familiar advertising request:

The one thing may be the position of a character, the direction of a glance, the visibility of a product, or the warmth of the light. In traditional visual-effects work, such a request might be tedious, expensive, or technically difficult. But the request is at least conceptually coherent. A scene exists. Its constituent parts exist. They can be identified and altered.

In a generated image, the constituent parts do not always exist in the same sense. The model has not built a room and then placed a person inside it. It has produced an image of a person in a room. The distinction is invisible until someone asks to move the person.

Then the room may change.

The furniture may change with it. The person may acquire a new face. The camera may rise, the window may narrow, the costume may mutate, and the light may begin arriving from an entirely different sun. The requested revision has been made, but so have several hundred unrequested ones.

AI is remarkably good at giving you another answer. It is less reliable at giving you the same answer, amended.

02 - The LocationsRural Britain, Statistically Recalled

This was one of the central problems on the Gigaclear commercial. The landscapes needed to resemble rural Britain rather than an idealised version of it. The references had to hold the right roads, stone, vegetation, weather and sense of lived-in place.

The locations had to feel plausible enough that a Gigaclear vehicle could enter them without appearing to have driven into another genre.

So the villages were generated, reviewed, rejected, adjusted, and generated again. The roads were too broad. The houses were too picturesque. The stone was wrong. The vegetation was wrong. The weather was too dramatic. The weather was not dramatic enough. Some images looked convincingly British but insufficiently rural. Others were rural in a way that suggested nobody had lived there since the invention of electricity.

Each new generation arrived quickly. This created the sensation of speed.

The distinction between speed and progress is important. Generative AI increases the number of attempts a production can make. It does not necessarily increase the rate at which the production reaches an approved result.

The archive fills. The schedule does not move.

Much of AI production is a process of selection and correction: generate, inspect, compare, discard, then generate again. A promising result may be difficult to recreate. A saved frame can reveal a problem only after it is viewed at full size. Clear versioning and a repeatable workflow matter because the archive grows quickly.

This is why the figure of 4,063 images should not be understood as a measure of productivity. It is closer to a geological record. Buried inside the folder are the discarded evolutionary branches of the commercial: places the project nearly went, characters it almost adopted, compositions that survived for a day and then disappeared.

Only a small portion of those images became usable material. A smaller portion became shots. Fewer survived the review process. Fewer still appeared in the finished film.

That ratio is not evidence that the system failed. It is evidence that the system encountered production.

03 - The CharacterThe Giganaut Problem

The most persistent encounter involved the Giganaut.

A conventional character pipeline is cumbersome but philosophically straightforward. A character is designed. The design is approved. A model is constructed. Materials are assigned. A skeleton is created. Once those tasks are completed, the character possesses a stable existence. It can enter a different room without becoming a different person. Its helmet does not reinterpret itself in response to the furniture.

Generative AI has a more flexible understanding of identity.

The Giganaut could be produced rapidly and attractively. He appeared in hangars, on stages, in domestic interiors and against rural landscapes. In isolation, many versions looked successful. Across the sequence, however, the costume, proportions and face did not remain consistent enough to represent one approved character.

The orange changed. The helmet changed. The proportions changed. Some versions became more astronaut and others more superhero. The general idea remained, but the approved character did not.

Every version preserved the general idea. Few preserved the specific character.

This is one of the more consequential limitations of current generative systems: they are highly capable of preserving semantic identity and much less dependable at preserving production identity. They understand that the subject is an orange-suited astronaut. They do not necessarily understand that it must be this orange-suited astronaut, with this helmet, these proportions, and the exact details that were approved three meetings ago.

A viewer may accept the first kind of consistency. A brand requires the second.

Eventually, the Giganaut was modelled, textured, rigged, and rendered in three dimensions. AI had helped locate the character. Traditional production methods made it possible for him to continue existing.

This was not a retreat from the new technology. It was an example of using it correctly.

AI is particularly effective at generating possible worlds. Three-dimensional production is effective at maintaining a world after it has been chosen. Compositing reconciles that world with the other elements in the frame. The useful pipeline was not a competition between tools; it assigned each tool the work it could control reliably.

This arrangement is less simple than the idea of a single prompt replacing a studio, but it creates a process that can be directed, revised and delivered.

04 - The EconomicsBeautiful, Cheap, and Wrong

Hollywood has a long history of greeting new technology as a labor-saving device. The language changes, but the promise remains consistent: the next system will allow the industry to produce more material, in less time, with fewer people, and with greater creative freedom.

Often, the technology does allow some of those things. It also creates new departments, new dependencies, new specialists, new failures, and new forms of work that were difficult to anticipate from the demonstration.

Digital cameras reduced the cost of film stock and increased the amount of footage that could be captured. They did not reduce the need to determine which footage mattered. Computer graphics made impossible images possible, then created an industry devoted to making those images look as though they had not been made by computers. Virtual production moved locations onto LED walls and then required large teams to construct, operate, synchronize, light, photograph, and repair the worlds on those walls.

AI follows the same pattern, but at a more disorienting speed. It makes the beginning of the process look so easy that the remainder can appear unnecessary.

Executives see a generated frame and mistake the frame for the production. Clients see an image arrive in seconds and assume that the final image should arrive shortly afterward. Software companies present a journey from "idea" to "execution" that consists mostly of an arrow.

Missing from the arrow is everything that happens after someone has an opinion.

A commercial is not expensive because pressing the record button is difficult. It is expensive because hundreds of decisions must be made to agree with one another for thirty seconds. The actor, product, performance, lens, lighting, set, environment, edit, sound, grade, graphics, legal requirements, and brand identity must all point in approximately the same direction.

AI can produce raw material for many of those decisions. It does not assume responsibility for their agreement.

The distinction is easy to miss because beauty announces itself immediately. Specificity tends to become visible only when it is absent.

A generated image may possess excellent composition, rich lighting, attractive production design, and the surface evidence of enormous expense. It may also contain the wrong product, the wrong person, the wrong architecture, or a camera position that makes it unusable in the sequence. It may be beautiful in precisely the way the project does not need.

Production attaches obligations to beauty. The woman must be in this room, wearing this outfit, holding this object. The light must arrive from this direction. The product must be legible but not unnaturally prominent. The scene must feel expensive but not inaccessible, futuristic but not cold, rural but not nostalgic, polished but still human.

A client rarely asks for the impossible in the language of impossibility. The request arrives as a series of modest adjustments.

Make the product cleaner. Keep the reflection. Lower the camera. Keep the face. Move her slightly left. Keep the room. Make it feel more premium. Keep it natural.

Each instruction is small. Together, they describe a system of control that the underlying generator may not possess.

The valuable person in an AI production is therefore not necessarily the person who can produce the most impressive initial image. It is the person who can diagnose why an impressive image is wrong and determine the least destructive way to repair it.

Sometimes the answer is another generation. Sometimes it is an inpainted region, a mask, a depth pass, a projection, a three-dimensional reconstruction, a paint fix, a change in the edit, or several hours spent rotoscoping a hand because the model has treated anatomy as an invitation to improvise.

Sometimes the answer is a conversation.

Clients frequently describe visual symptoms rather than technical causes. They say that an image feels artificial, that a character lacks presence, or that a shot does not feel premium. The actual problem may be a mismatch in lens perspective, insufficient contact shadow, overly sharp background detail, inconsistent motion, incorrect skin response, or a subject who appears to occupy a different atmosphere from the room.

The model cannot reliably diagnose that problem, because the model does not understand the purpose of the shot. It does not know what was approved, what the next shot contains, what the brand is trying to communicate, or why a technically beautiful result has made everyone on the call uncomfortable.

05 - The LeverageWhere AI Actually Earned Its Place

On the Gigaclear project, the most productive use of AI did not involve asking it to create a complete commercial. It involved identifying particular production problems whose conventional solutions would have consumed disproportionate time and labor.

One sequence required a crowd of roughly one hundred background characters inside a hangar. Traditionally, this might have involved casting and photographing many people, extracting them, preparing them as assets, arranging them in depth, matching the lighting, and rendering or compositing the final scene. The task was possible. It was also the sort of task capable of converting several production days into a blur of masks, files, and tiny human silhouettes.

Instead, a pipeline was built. Characters were generated, converted into usable assets, brought into Cinema 4D, and distributed through the scene automatically. The crowd was completed by one artist in roughly a day.

This is the kind of reduction in labor that AI can genuinely provide. The technology did not determine what the shot should communicate. It did not replace the director's judgment or the artist's eye. It removed a repetitive bottleneck that might otherwise have made the shot impractical.

The valuable system is not one that generates indiscriminately. It is one that solves a defined production problem without damaging the intention surrounding it.

That requires engineering, but it also requires taste. A pipeline must know which variables can be automated, which must remain under human control, and where a shortcut will produce consequences elsewhere in the sequence. An elegant automation that produces the wrong crowd is useless. A brilliant visual decision that cannot be repeated across one hundred characters is unaffordable.

The emerging profession sits between those failures.

The person performing it may be called an AI workflow specialist, a creative technologist, a generative-production lead, or some other title that gives an established form of judgment the appearance of having been recently invented. The title matters less than the ability to move among disciplines: generation, three-dimensional production, compositing, editing, scripting, versioning, review systems, and client communication.

The most important skill will remain familiar. It is the ability to look at a shot and understand why it does not work.

06 - The CraftThe Seam

There were real actors in the Gigaclear commercial. This fact tends to complicate the cleaner stories told about synthetic production.

The actors were not present because AI had failed to replace them. They were present because replacement was not the purpose of the project. People remain unusually capable of being watched by other people. A human performance contains kinds of information that are difficult to itemize: hesitation, weight, timing, self-consciousness, the minute changes that occur when one person reacts to another. A face does not merely occupy a frame. It creates a reason to keep looking at it.

The generated environments allowed those performances to exist inside locations and circumstances that would have been expensive or impractical to photograph. AI replaced logistics, extensions, background populations, and some forms of physical construction. It did not replace the need for a human event at the centre of the image.

Afterward came the older work.

The generated locations had to be matched to photographed material. Light had to appear to belong to one world. Perspectives had to agree. Edges had to be repaired. Faces, costumes, environments, and three-dimensional elements had to be integrated. Motion had to remain stable. The grade had to suppress the small disagreements that cause an audience to sense, without necessarily understanding, that something is wrong.

Every major image-making technology eventually encounters this stage. Green screen encounters it. Computer graphics encounters it. Virtual production encounters it.

It is the seam.

On Gigaclear, the seam was most visible where a photographed actor entered a generated location. The task was not to make either element impressive on its own. It was to make the contact shadow, perspective, grain, colour and movement agree long enough that the audience read one place, not two layers.

Generated material flickers, drifts, mutates, forgets objects, changes textures, and solves each frame with slightly different assumptions. An individual frame may be persuasive while the sequence surrounding it remains unstable.

Cinema is not a collection of successful frames. It is memory across frames.

An audience remembers where an object was, how a face looked, which direction the light was travelling, and what kind of world the previous shot established. The audience may not consciously document these facts, but it notices when the film forgets them.

07 - The BudgetThe Revision Path

This is also why the client does not truly pay for the prompt.

The prompt creates possibility. The revision creates the product.

During the first review, everyone responds to the general idea. During the second, they become specific. By the third, the brand guidelines have re-entered the discussion. Later, legal may notice the product representation, the director may miss the energy of an earlier version, and someone will ask for the emotional quality of version two with the composition of version five and the approved product from version seven.

The team will be told that the project is almost there.

"Almost there" is not a measurable distance. In production, it can be longer than the entire journey that preceded it.

The apparent speed of the first generation creates a dangerous expectation that each later step should be equally fast. But the first generation was asked to invent. Every subsequent version is asked to obey an expanding history of decisions.

The project becomes harder as it becomes more defined.

The budget is therefore not determined primarily by the length of the prompt, the price of a generation, or the name of the model. It is determined by the revision path. How many shots must agree with one another? How consistent must the characters remain? How many stakeholders can request changes? Which elements have been approved and therefore cannot move? How much of the work must withstand agency, director, brand, and legal review?

Those questions describe the actual scale of an AI production.

A demonstration can contain five mistakes and still be admired because the achievement lies in its existence. A commercial can contain one conspicuous mistake and be rejected because the achievement lies in its reliability.

08 - The LessonWhat the Film Proved

The Gigaclear film worked because it was never treated as a magic trick. A director established the intention before the tools arrived. AI was used where AI created leverage. Three-dimensional systems were used where objects and characters needed to persist. Traditional visual effects were used where separate elements needed to become one image. Human performances remained where human performances mattered.

The production succeeded not by finding a single technology capable of doing everything, but by preventing each technology from doing the things at which it was unreliable.

This is the less glamorous future of artificial intelligence in commercial filmmaking, and probably the more consequential one.

It will not consist mainly of a person entering a paragraph and receiving a finished advertisement. Real production will involve smaller teams doing larger work through interconnected systems: generative models, asset libraries, reference controls, three-dimensional scenes, compositing pipelines, automation scripts, review tools, and artists capable of moving between all of them.

The models will improve. Characters will become more consistent. Editing and revision controls will become more precise. This will not make taste less important.

As generation becomes cheaper, the supply of attractive imagery will increase. More companies will be able to produce work that looks, at first glance, expensive. The advantage will move away from the ability to create visual material and toward the ability to control, select, structure, and finish it.

The basic output will become cheaper. Judgment will become more valuable.

AI has made the image appear to be the easy part. In one sense, it is. A person can now produce in seconds an image that would once have required a set, a crew, a camera, and a considerable budget.

But production was never merely the creation of an image. It was making the image specific. Making it repeat. Making it survive contact with other images. Making it support a performance. Making it remain coherent after the client, agency, director, legal team, editor, and brand have each asked something of it.

It was making the image hold up.

Near the end of a production, the number of possibilities begins to contract. Thousands of images have become dozens of shots. Dozens of shots have become a sequence. The sequence has acquired timing, meaning, and an internal history. By then, almost everything has been decided.

That is usually when someone asks to change one thing.

The request sounds small because only one thing has been named. But every part of a finished image is connected to every other part: the composition to the lighting, the lighting to the environment, the environment to the character, the character to the edit, the edit to the story.

Changing one thing while preserving everything else is not a minor function of production.

It is production.

Artificial intelligence can now generate the world before the sentence has finished echoing around the room. The difficult part is persuading the world to remain where it was placed.

Frequently Asked Questions

How many images does an AI commercial actually take?

On the Gigaclear spot the archive passed four thousand files, most of them near-duplicates, the visible trace of a real process: generate, compare, reject, regenerate.

Where did AI genuinely earn its place on the Gigaclear commercial?

In the locations, environments, and a 100-character crowd. Work that would otherwise mean scouting, travel, sets, and casting.

Why was the hero character built in 3D instead of generated?

Because raw generation couldn't hold the Giganaut's form across shots. A rigged 3D model gave the control that consistency demanded.

What is the biggest hidden cost in AI production?

The revision path. The first beautiful image is almost never the job; the budget lives in the iterations that make it right and on-brand.