The Aoki Bud Light shoot in 2018 is the shoot I keep going back to. Not because it was the biggest. Bigger shoots happened on both sides of it. I go back to it because it is the cleanest case I have of a project that ran a full traditional pipeline once, in real production, with a real budget, and which I have now rebuilt twice in 2026 in AI tools, once for myself and once as a teaching pass for a client who wanted to understand what changed. The frame by frame comparison is the most honest version of the answer to the question I get asked the most. What did AI actually change.
The 2018 shoot, by the numbers
The brief came through Common Good in Toronto. The client was Anheuser-Busch InBev through Bud Light. The talent was Steve Aoki. The deliverable was a sub-series inside the Bud Light Living master campaign, with cuts for broadcast, social, and behind the scenes. The locations were Las Vegas for the stage, Toronto for the green room, and exterior plate in a third city which I will not name because that was a closed lot.
Pre-vis took six weeks. I will break that down because the number on its own does not communicate the texture. Week one was the script lock and the agency revisions. Week two was the storyboard pass with the lead board artist. Week three was the animatic, hand assembled with library music and rough motion. Week four was the body type reference shoot, which sounds clinical and was, because the agency needed to approve the silhouette of every dancer in the venue scenes before casting could even start its real work. Week five was the recut animatic with the new references, plus the second agency revision. Week six was the lockdown and the cost report.
The agency, the client, the production company, and the director all signed off on a document that took six weeks and roughly forty thousand dollars in pre-vis spend to produce. The shoot then ran four days of principal across two countries. The edit ran nine weeks. The total cycle from brief to broadcast cut was four months and change.
The 2026 reconstruction
I rebuilt the pre-vis in May of 2026 using the hybrid pipeline I wrote about at length elsewhere on this site. Veo for the talent reference and the green room scenes. Sora for the stage moves and the establishing exterior. Kling for one specific dream-logic insert that the original cut did not have but that the reconstruction does because the tools made it cheap enough to add.
Total time from script note to a watchable rough cut was thirty minutes. Not thirty hours. Thirty minutes. I am being precise because the number sounds rhetorical. It is not. The script note was at 14:02. The rough cut existed at 14:34. I have the file timestamps.
The body type reference shoot, which in 2018 took a full week and cost three thousand dollars in talent and lighting, was now twelve minutes of model output. The agency would see the same approval surface. The casting team would still cast real humans against the AI silhouette. The function of the reference does not change. The labor of producing it collapses.
The agency proof-of-concept cutaways, which in 2018 were a separate sub-shoot stitched in during the post window, now exist inside the pre-vis itself. The agency sees them on day one. The client sees them on day one. If the client hates them, you cut them at zero cost. In 2018, if the client hated them, you had already shot them.
The storyboard pass, which in 2018 cost a board artist a week of work, was now five hours of model curation. The actual visual quality of the boards is higher in 2026 because the model can produce a finished frame rather than a stylized indication. The board artist as a role does not disappear. The role becomes the curator of model output rather than the maker of pen and paper indications. Some board artists made that transition easily. Others did not.
What stayed the same
The director still picked every shot. That has not changed. Every frame that survived the cut in 2018 survived because I picked it. Every frame that survived in the 2026 reconstruction survived because I picked it. The model produced a thousand frames a day. The director picked the fifty that mattered. That is the same labor it has always been. The unit of measurement is the eye, not the file.
The cast still mattered. In 2018, Aoki was Aoki because Aoki has been Aoki for two decades. No model output can replicate the specific gravity of a real performer who has earned a specific reading from an audience over a real career. In the 2026 reconstruction, the Aoki frames are an approximation. They are good enough to plan a shoot. They are not the deliverable. The deliverable is still a real human in a real frame.
The cut still mattered. The most important shots in the 2018 master cut were the ones I held a beat longer than the agency wanted. Those shots survive the 2026 reconstruction because I held them again, on instinct, the same instinct. The tool does not know which shots to hold. The director knows.
What changed
The cost of iteration changed. That is the only honest answer. Forty thousand dollars and six weeks for the 2018 pre-vis. Roughly two hundred dollars in inference credits and thirty minutes for the 2026 reconstruction. The ratio is not subtle. It is two hundred to one in dollars and roughly two thousand to one in time.
The director still made the shot. The AI made the iteration possible. That is the sentence I have settled on after rebuilding this project twice. The director still made the shot. The AI made the iteration possible.
Amos LeBlanc, the older byline spelling, was the EP credit on the 2018 master campaign through Common Good Toronto. The 2026 reconstruction sits in the working archive at amosleblanc.com/ai and is the cleanest teaching case I have for what the new pipeline actually does. It does not replace the director. It returns the director to the question they were supposed to be answering all along, which is what the shot is for.