How an AI ad is made, from character sheet to cutdowns

There is no shoot, so the decisions move earlier

A conventional commercial spends its money around the shoot: the set, the crew, the travel, and the reshoot when something goes wrong. An AI ad has none of that. Run AI Studios does not run shoots, crews or sets; every frame is generated and then edited. The decisions that used to be made on set therefore move to the start of the job, where they are cheapest to change.

Why the character sheet comes first

Viewers forgive a lot in an ad. They do not forgive a lead whose face changes between two shots, and that drift is the best-known failure of generated video.

So before a single shot is made, every lead is locked as a reusable identity: face, wardrobe states, age and era. The world around them follows the same rules: “Locations, props, creatures and brand assets follow the same rules — one product, one pack shot, one world, held across an entire campaign of assets.”

The character sheet is that lock, written down: the canonical reference every later frame is compared with. The comparison is a formal gate of its own, the identity gate, so a face that has drifted is caught before it reaches the cut.

The six stages

Run AI Studios makes an ad in six stages, and a person reviews the output of every one. The pipeline is on the Studios page; here it is step by step:

1. Develop — Script engine. Competing scripts, independently judged; the winner becomes a per-scene shot list.
2. Look-dev — Locked film look. One grade, one light logic, film-stock texture — identical in every frame.
3. Frames — Keyframe first. Every key still is human-approved before any motion spend is committed.
4. Motion — Cinematic shots. Physics-true movement, single camera moves, one beat per cut.
5. Sound — Score & design. Original score and effects planned per second, blind-QC’d before the mix.
6. Finish — Conform & master. Cut, mix, grade, titles — cinema master and broadcast versions delivered together.

Keyframe-first economics: we iterate where mistakes are cheap and commit only when the frame is right. That discipline is why the quality holds and the cost doesn’t balloon.

Stills before motion: where the budget is protected

The most important habit in the pipeline is the order of the frames and motion stages. Every key moment of the ad is made as a still first, and a person approves it before any motion is generated. A still is quick to redo; a moving shot costs far more to throw away.

This is also where the client directs. In a typical sprint, the second day is the frames: “You approve stills of every key moment. Total creative control, zero committed risk.”

The gates between a generated shot and the brand

Generated footage fails in recognisable ways: a face that drifts, movement that breaks physics, audio that sounds synthetic, an edit that runs long. Each failure has a gate of its own, and nothing moves on until it passes. There are six:

• Frame gate. Every keyframe reviewed by human eyes before animation.
• Identity gate. Every face checked against its canonical character.
• Motion gate. Every clip reviewed frame-by-frame for physics and artifacts.
• Audio gate. Blind sound QC — defective or synthetic-sounding audio is rejected.
• Edit-math gate. Runtime, pacing, title discipline, restraint in the mix — verified.
• Final watch. Complete watch-through with fresh eyes before anything ships.

Some rules sit upstream of the gates, in how each shot is written, because they stop the “AI look” before it is generated: one physical action per cut, faces live in stillness, motivated single-source light and two loud moments maximum.

Some subjects are harder than others. The notes on a food spot in the archive, Sandwich Brand Spot, put it plainly: texture and weight give generated food away first, so the close-ups are where it has to hold. Jewellery is judged on detail in the same way, which is why the macro shots carry Bridal Jewellery Film.

From one master to every cutdown

An ad rarely runs in one shape. A campaign needs a widescreen master, vertical and square versions, shorter cuts and often other languages. In this pipeline they are all struck from the one approved master, as platform-native cutdowns, aspect ratios and language versions, rather than made again.

Cutdowns are also where sound gets tested. The archive holds one saree spot exported twice, Saree Campaign — Voiceover Cut and Saree Campaign — Clean Cut: the same edit with and without narration. A feed plays muted by default, so an ad that only lands with a voice track only lands for part of its audience.

ConceptAI Ads

Saree Campaign — Voiceover Cut

The sound-on half of a matched pair: one apparel spot exported twice, so the same edit can be judged with narration and without it.

Open the artifact

ConceptAI Ads

Saree Campaign — Clean Cut

The same apparel spot with the narration removed — the version that has to carry the offer on picture alone.

Open the artifact

What the finished work looks like

The work archive holds 39 AI ads and 5 UGC-style pieces, each on its own page. Three to start with:

ConceptAI Ads

Sandwich Brand Spot

A 27-second food spot: a sandwich built ingredient by ingredient in a warm kitchen, then the first bite.

Open the artifact

ConceptAI Ads

Bridal Jewellery Film

An 18-second vertical jewellery film: a bride alone in a field of white cones, then the pieces in close-up.

Open the artifact

ConceptUGC-Style

Street Gossip Clip

A 34-second vertical street clip in the gossip format, captioned word by word: “Weren’t you in Tokyo?”

Open the artifact

What we do not do

• We do not run shoots. There are no crews, sets or shoot days in this pipeline.
• We do not name the models we use. The models are commodities — the pipeline around them is ours. We run the strongest tool for each job and swap as the frontier moves. Single-model studios ship what their model can do; we ship what the brief needs.
• We do not use anyone’s face or voice without consent. Any real person’s likeness or voice runs on written, revocable consent with secure storage — never scraped, never assumed.
• We do not publish prices. Every engagement is priced in a written proposal for a defined scope.

Questions

How long does an AI ad take to make?

A typical short-form sprint runs from a brief on Monday to a premiere on Friday, and larger programs run in weeks. Measured on our own production, a typical teaser runs ~48h from brief to master.

Is any part of an AI ad shot on camera?

No. Run AI Studios does not run shoots, crews or sets for an ad. The frames are generated from the script, the character sheets and your references, then animated, scored and edited.

Can the same character appear across a whole campaign?

Yes. That is what the character sheet is for: each lead is kept as a reusable identity (face, wardrobe states, age and era), and every frame is checked against it at the identity gate before it moves on.

Which AI models does Run AI use?

The models are commodities — the pipeline around them is ours. We run the strongest tool for each job and swap as the frontier moves. Single-model studios ship what their model can do; we ship what the brief needs.