How we made the Bugatti Tourbillon ad with AI (in 3 weeks).

Every AI video model is trained on what already exists on the internet. The Bugatti Tourbillon had launched maybe a month before we started. So when we typed "Bugatti Tourbillon" into any generator, we got Chirons, Veyrons, and weird hybrids. The model had no real reference.

Akanksh and I had two choices: wait for the model to catch up (it wouldn't, in 3 weeks), or build the look from scratch. We built it from scratch.

A character sheet in AI work is the same thing it is in animation. It's a multi-angle reference of one subject so the model knows what to render every time you ask for it. For the Tourbillon, we built a sheet across seven specific angles.

Each angle took 15 to 25 generation passes to get right. We used Bugatti's official press shots as anchors, then iterated until the proportions, body lines, and signature details were dead consistent.

Total images burned to lock the car: about 350. Total kept for the character sheet: 8.

AI models are trained on celebrity images, so they'll generate "a Kendall Jenner lookalike" — but they break down the moment you ask for specific angles, expressions, or lighting conditions.

We built her sheet around six anchor frames:

• Driving position — hands at 10 and 2, eyes on the road.
• Side profile in motion — for cutaways during the drive sequence.
• Smiling at camera — low angle, hero shot.
• Behind the wheel — interior, framed against the dashboard.
• Walking away from the car — full body, scale reference.
• Close-up of eyes and mouth — the brand moment.

Took us about 250 images to lock her character sheet. The hardest part: getting consistent eye colour, jaw structure, and hair behaviour across angles.

The model knows Iceland. Black sand beaches, basalt columns, glacial lagoons, the works. We pulled together a location moodboard from real Iceland photography (Diamond Beach, Stokksnes, Vík), then generated environments to match. Four locations made it into the film:

• Reynisfjara black sand beach — driving sequence.
• Stokksnes with Vestrahorn mountain — hero shot.
• Diamond Beach — Kendall walking up to the car.
• A long basalt road through fog — the closing shot.

Iceland took us about 200 environment generations. Less than the car or Kendall because the model already had the data.

Once we had the Bugatti sheet, the Kendall sheet, and the Iceland environments, we started building the shots. Dropping a character into an environment usually breaks one or the other — so we ran a stacked pipeline:

• Reference image conditioning — feed the model the car sheet + Iceland environment together.
• Multi-image input — Seedance 2.0 lets you reference up to 12 files at once.
• Img-to-vid generation for motion — Veo 3.1 and Kling 3.2 for the actual movement.
• Manual frame fixes — Photoshop and After Effects where the AI broke.

AI doesn't edit your film. We do. Akanksh sat for five days in the edit timeline — colour grading the AI footage so it matched a unified look. Adding camera shake and grain so it felt less synthetic. Cutting on motion to hide the model breaks.

Sound design used real Bugatti engine audio from the manufacturer's reveal video, layered with synthesized low-end. Iceland ambient (wind, distant waves) underneath. A custom score on top.

Without the edit, AI footage looks like AI footage. With the edit, it looks like a Bugatti commercial.