How an AI commercial gets made: the seven stages, in order
Making the pictures is one stage of seven. This is the order the work runs in, what you are shown along the way, and the stage that decides whether shot two matches shot one.
Early on an AI job a test image arrives, and it looks finished. For a moment the whole commercial seems solved. I understand the appeal, because I felt it too. But a commercial has to get through your notes, a brand check, a TV channel's technical rules and a delivery date, and one good image tells you nothing about any of them.
The work runs in seven stages, and making the pictures is the fourth. The stage that decides whether a face or a product looks the same in every shot comes before it: whatever has to stay the same is built first, and the pictures are made on top.
What follows is the order I work in. By the end you will know what to ask anyone who offers you AI production, and at which stages you see the work and approve it.
01 · Concept & Direction
Deciding what the film is before a single frame exists
Every later stage inherits its limits from this one, so it is worth being slow here. Concept work fixes what cannot change: what the brand is selling, what must stay literally true about the product, the tone, and which shots carry the argument. On the Gigaclear commercial three facts were set first. Real actors would be placed in generated environments. A crowd of roughly a hundred had to read as one community. And the whole thing had to hold up on national TV. All of that was decided in words and references before any model was used. When the tool sets the direction instead, a striking test frame can end up dictating the concept, and that usually means expensive corrections later.
You see: the brief written back as a short list of fixed facts, with references. You sign off: what cannot change, before any model is used.
02 · Previs
Where AI helps first, and where it stops
AI helps here first. With generative models I can build a previs, a full visual treatment of the film, and it can arrive far sooner than a conventional treatment. You can judge the film while it is still only a plan.
A previs is a preview. It looks good and it is not production-ready: it does not yet have the consistency or the exactness a TV deliverable needs. It proves the idea and gets everyone agreeing on it. Then it becomes the brief for the stages that follow, which exist to make it real and repeatable.
You see: the previs, a full visual treatment of the film. You sign off: mood, light, framing and overall feel, before the larger budget is committed.
03 · 3D
Building the control before the generation
This stage comes before generation on purpose. Once the previs is approved, I build the parts that have to stay exactly themselves as real 3D, in Cinema 4D. For Gigaclear, the hundred-character crowd was rebuilt as geometry, so each character could be posed, lit and moved. Products, key environments and any shot that needs a precise, repeatable camera move are built or laid out here too.
A 3D scene can be directed: the same camera, the same layout and the same identity from every angle. The generation is wrapped around it, so the model adds its look without redrawing the layout and the characters in every frame. I explain how characters and products are kept consistent in a separate article.
04 · Generation
Generating on top of the 3D
Now the model does what it is good at, against a clear target. A shot generated from a sentence alone may not match the one before it, so I generate on top of the 3D. The rendered layout, depth and geometry steer the generation. The 3D holds the composition and the identity, and the model supplies texture, light and the photographic look. Used this way, generation adds atmosphere and detail without the drift that makes prompt-only work unreliable when something has to look the same twice.
The generation itself runs on node workflows instead of prompts, and how the model for each shot is chosen is a separate decision.
05 · Visual Effects
Cleanup, stitching, and the major fixes
No generation arrives finished. Slow a clip down and a brick wall shimmers or a jacket weave crawls, because the model redraws each surface in every frame. So the first job is temporal cleanup: stabilise, deflicker, rebuild a clean plate, and replace a surface outright where the shot depends on it. Real actors are then composited into the generated world, with edges, contact shadows and depth matched so they sit in it. Anything added to a moving shot, such as a logo, a product or a set extension, is tracked so it shares the plate's motion. Hands, logos and text that came out wrong are repaired with paint and tracking, or replaced with real type and clean graphics.
This is the stage I came from, and it has its own article: the VFX work an AI shot still needs.
06 · Grade & Finishing
Making seven stages read as one film
Once the frames hold, the film has to feel like one film. The grade pulls the generated, rendered and filmed material into one palette, and corrects the drift that makes a warm shot and a cold shot look as if they belong to different afternoons. Without it the sequence can be technically clean and still feel like several films. Sound design and the final edit carry as much of the emotion as the picture.
07 · Delivery
The unglamorous stage that decides whether it airs
A film that cannot be delivered to spec does not exist as far as a TV channel is concerned. This stage is versions and compliance:
- TV master to the channel's exact technical spec: resolution, frame rate, audio loudness, safe areas, colour space and legal levels.
- Social and online versions in their own ratios and durations, reframed rather than crudely cropped.
- Textless and versioned masters so the spot can be localised or amended without a rebuild.
A pipeline that treats generation as the finish line can arrive here with one beautiful file and no way to conform it. The work reaches delivery intact because the control was built in stage three, before the model generated a frame.
Frequently Asked Questions
What are the stages of an AI production pipeline?
Concept and direction, previs, 3D, generation, VFX, grade, and delivery. Each stage exists because a commercial has to be repeatable and ready to air, and a striking single image is neither.
Why build in 3D before generating with AI?
Raw generation can't hold a shape or a camera move across shots. Building the scene in 3D first gives a control layer to generate on top of, so the result stays consistent and directable.
Is AI previsualization production-ready?
No. Previs is where AI helps first, because it shows look and blocking quickly. It isn't final quality. Treat it as the plan for the stages that follow.
Where does traditional VFX fit into an AI pipeline?
At stage five, after generation: temporal cleanup, compositing real actors into generated environments, tracking logos and products into moving shots, and repairing hands and text so every shot can survive an edit and delivery.
What is the hardest stage to get right?
Delivery. Specs, formats and QC decide whether a spot airs, and a pipeline that stops at generation has no way to conform to them.
Send the script or the deck. Within a day you get a written reply that marks it up against these seven stages and shows which shots carry the risk. That first read is free, and if AI is not a fit I will say so.
Send the brief → See how the Gigaclear commercial was made →