Final Pixel: Five Times AI Already Made the Cut
Everyone’s still arguing about whether AI belongs in filmmaking. Meanwhile, it’s already been in five things you’ve watched.
Not concept art. Not a prompt-generated trailer nobody saw. The actual final pixel — the frame that shipped, that you paid for, that’s sitting in your watch history right now.
Here’s where it already happened, and what the pattern tells us.
The invisible wins
Dune: Part Two needed roughly 1,000 shots of Fremen eyes glowing that distinctive blue. On the first film, VFX supervisor Paul Lambert’s team hand-rotoscoped every one — outlining pupil, iris, and sclera frame by frame. For the sequel, with four times the Fremen, that approach wasn’t viable. So Lambert trained a machine learning model on 30,000 eyes from the first film’s footage and let it generate the mattes automatically. It won Lambert his fourth Oscar.
Marvel’s Ant-Man and the Wasp: Quantumania used something called Vanity AI, built by a Toronto studio called MARZ, for cosmetic touch-ups on Kathryn Newton’s Cassie Lang — smoothing a wig line, cleaning up skin under harsh lighting. The same tool has quietly worked on Spider-Man: No Way Home and Stranger Things. Nobody noticed. Nobody was supposed to.
Neither of these made headlines. That’s the point. When AI handles the drudge work invisibly and the artist stays in control of the outcome, audiences don’t just tolerate it — they don’t know it happened.
The one that got it right
In April 2025, Netflix’s Argentine sci-fi series The Eternaut needed a building to collapse across Buenos Aires. It’s a two-second shot, buried in episode six. Traditional VFX would have blown the show’s budget. So Netflix’s internal team, Eyeline Studios, built it with generative AI instead.
Here’s what made this one different: Netflix didn’t wait to get caught. Co-CEO Ted Sarandos announced it himself, on an earnings call, unprompted. “The creators were thrilled with the result. We were thrilled with the result. And more importantly, the audience was thrilled with the result,” he said. Ten times faster than traditional workflows. A shot that simply wouldn’t have existed otherwise.
The show got a 95% critics score on Rotten Tomatoes and a season two renewal. Nobody boycotted it.
The ones that got caught
Compare that to Secret Invasion. Marvel’s title sequence used AI-generated imagery for its shifting, uncanny Skrull visuals — a legitimate creative choice for a show about shapeshifters. The problem wasn’t the tool. It was the discovery. A journalist broke the story, the director admitted he didn’t really understand how the AI worked, and the reveal landed in the middle of the WGA strike, which made it a lightning rod instead of a footnote.
Indiana Jones and the Dial of Destiny used a genuinely impressive system called ILM FaceSwap — machine learning combined with over 100 artists and three years of work — to make an 80-year-old Harrison Ford look 40 again for the film’s opening. That one mostly avoided backlash, probably because ILM was upfront about the effort and artistry behind it from the start.
The pattern
Put these side by side and the difference isn’t the technology. Dune’s eye replacement and The Eternaut’s building collapse used comparable machine learning techniques to Secret Invasion’s title sequence. What separated a warm reception from a firestorm was disclosure and intent.
Tell people what you did and why, before they find out on their own, and it reads as craft. Get caught, or seem confused about your own process, and it reads as a shortcut.
That’s the same lesson from the A24 and Google DeepMind deal I wrote about a few weeks ago. The structure of disclosure is becoming the actual product decision, not an afterthought bolted on after backlash.
We’re long past the debate over whether AI belongs in the pipeline. Five real productions, going back to 2023, already answered that question with their final cut. The debate that actually matters now is who tells the audience, and when.