500 AI shots on prime-time CBS

Harlan Coben's Final Twist returned to CBS on 5 October 2026 with ten real crime scenes rebuilt from almost nothing. The studio behind them, Promise, operating as The Generation Company, delivered more than 500 AI-generated shots for the season. It is the best-documented piece of delivered AI work in entertainment this year, and almost nothing about how it was made resembles what a brand marketer imagines when a deck says "AI production".

Nem Perez, Head of Studio at The Generation Company, framed the brief as a question: "How do you use GenAI to reconstruct a crime scene from only a handful of archival images and make it feel real?" That is a reconstruction problem, not a generation problem. The scenes existed. Most of them no longer do, or were never photographed from the angles the show needed. A handful of stills is the entire reference package.

“How do you use GenAI to reconstruct a crime scene from only a handful of archival images and make it feel real?”

The pipeline

Four things were used together, per Parade's reporting on 6 October 2026.

• Generative AI, for material that had to be invented because no reference exists.
• 3D, for the geometry that has to stay consistent when the camera moves.
• Gaussian splats, which turn a small set of photographs into a navigable volume rather than a flat plate.
• Unreal Engine, as the environment everything is assembled and rendered in.

Then the part that decides whether any of it reads as real. Coben was shot on a volume stage in Los Angeles against an LED wall, with camera-tracked parallax. The background moves correctly relative to the lens. The host is lit by the scene he is standing in. A human presenter, photographed properly, in front of an environment that behaves like a place.

That is a hybrid pipeline and it has a specific implication for cost. Every element on that list is a craft discipline with its own specialists, its own review cycle and its own failure modes. Generative AI removed the need to build or find ten locations. It did not remove the need for a stage, a crew, a lighting plan or a tracked camera.

You cannot prompt your way to the ends

Dave Clark, co-founder and chief creative officer of Promise, says it more bluntly than we would: "You can't prompt your way to the ends." Forbes frames the real technical problem as continuity of look. Holding one consistent look across a twenty-minute runtime is a different order of difficulty from producing one striking shot. Anyone can get a beautiful six seconds. The job is the hundredth shot matching the first.

Clark raises a second problem that gets almost no attention in vendor marketing, and it is the one that actually governs whether a studio can take a commission. How does a director communicate a vision to a team of a hundred people? Prompting is a private act. Production is not. A pipeline that only works when one talented person is driving it is a showreel, not a studio. Everything that makes a shot repeatable by someone else, references, lookbooks, shot specs, naming, versioning, review, is the same craft infrastructure film production has always needed.

Who Promise is

Backing came from a16z and Peter Chernin's North Road, with the Google AI Futures Fund and Crossbeam joining in May 2025. The slate as of July 2026 ran to 17 projects with 6 in production, including Hardcore 94, a four-part adult animation with its pilot in production, and Tuning In, an animated feature in pre-production.

The more revealing move was buying a school. Promise acquired the AI film school Curious Refuge, founded by Caleb and Shelby Ward, which has taught around 10,000 students across 170 countries in 11 languages. The number that should interest anyone staffing a project: 95% of current students are entertainment or advertising professionals. The talent constraint in AI production is not people who can use the tools. It is people who can use the tools and have also finished a job before.

The same bet, in Shanghai

iQIYI opened what it describes as China's first AI film studio in Shanghai on 2 November 2026, built on the same principle from the other direction. Shoot the performance live, then change the sets, lighting, costume and props in post. VP Liang ZHU put the reasoning in one line: "An actor's performance is the core of any film or series."

Two studios, two continents, two unrelated production cultures, and the same architecture. Capture the human. Generate the world. The division is not arbitrary. Performance is the thing audiences are best at reading and models are worst at sustaining. Environments are the thing that costs the most to build and that audiences inspect least.

What this changes in a brand brief

Most brand film budgets are eaten by things this pipeline genuinely compresses. Location scouting and permits. Travel for a crew. Set build for a world that exists for nine seconds. Reshoots because the weather turned. Those are real savings and they are available now.

What does not compress is the part of the schedule that decides whether the film works. Casting a face the audience believes. Lighting it. Directing a performance. Agreeing a look and holding it across every shot in the cut. On Final Twist the generated material carried the locations and a camera-tracked stage carried the host, and the ten crime scenes read as places because both halves were built properly.

If a studio pitches you a world and has nothing to say about how the human in it gets shot, the pipeline is half-built. Ask where the stage is.