Sixty-three to one

A 95-minute AI-generated feature film screened at the 79th Cannes Film Festival today. Higgsfield AI and ByteDance's Volcengine brought Hell Grind to the Croisette, billing it as "the world's first ever AI feature film." The trailer dropped Wednesday. The screening happened Thursday. The headlines wrote themselves.

The production data is more interesting than the headline.

Fifteen people. Fourteen days. Five hundred thousand dollars, four hundred thousand of which went to compute. Two-time Cannes Official Programme filmmaker Adilkhan Yerzhanov co-wrote the script. The team included professional directors, DPs, and editors. The core generation model was Seedance 2.0. They generated over 16,000 video clips. They selected 253 final shots for the first episode.

Sixteen thousand clips. Two hundred and fifty-three shots. Sixty-three generated clips for every one that made the cut. A 1.6% acceptance rate.

That ratio is the story. Not the headline.

The rejection is the filmmaking

Kubrick shot seventy takes. Fincher shoots forty. The number is not the obsession. The number is the search. Each take is a question: is this what I meant? The first take is information. The tenth is refinement. The fortieth is the filmmaker closing the distance between what they imagined and what the footage returned.

Hell Grind's team generated sixty-three clips for every shot that survived. That is not prompting. That is filmmaking wearing different equipment. The generation model produced the raw material. The fifteen filmmakers decided what to keep. The model had no opinion about which of its sixty-three attempts was best. It generated all of them with equal indifference. The humans in the room watched all of them, compared all of them, and threw away 98.4% of them.

The 15,747 rejected clips are where the taste lived.

Adil Alimhjanov, the production lead, told Variety something worth remembering: "You can't go into AI and say, 'Make me a 95-minute cool video.'" That sentence is the quiet correction buried beneath every "world's first" headline. The model cannot make the film. The model can produce sixty-three candidates for a single shot. The filmmaker decides which candidate earned the cut.

Two rooms at the same festival

The same week Hell Grind screened, Nikola Todorovic from Wonder Dynamics was at the Cannes Film Market making a sharper version of the same argument from a completely different direction.

Todorovic has spent a decade building AI tools that accelerate the VFX pipeline. Motion capture without suits. Lighting and compositing driven by machine learning. The Russo brothers used the technology on The Electric State. Autodesk acquired the company in 2024. Animation studios report going from thirty seconds of output per day to three and a half minutes. Production companies of five to seven people are doing things they never could before.

Then he said the thing that matters: "We're not in the business of 'let's prompt a movie.' I'm not a believer that you can prompt a movie, prompt a performance, prompt a camera move."

Todorovic builds AI that slots into the existing production pipeline. It handles the technical burden. It does not handle the creative decisions. The filmmaker still picks up the camera, blocks the scene, directs the performance, frames the shot. The AI removes the drudgery of rotoscoping, compositing, wire removal. The craft stays in human hands because the tools were designed to leave it there.

Hell Grind sits in a different room. The generation model produced everything. No camera. No set. No physical light source. No actor in front of a lens. But the production team still exercised creative judgment at every stage: scriptwriting, shot design, clip selection, editing, sound design. The creative decisions survived because the people making them had vocabulary.

Guillermo del Toro, presenting a 4K restoration of Pan's Labyrinth at the same festival, diagnosed the problem that connects both rooms. "They're trying to pass five types of things under one single name," he told The Hollywood Reporter. "What are you talking about? Are you talking about a tracking program? Are you talking about rotoscope? Or are you talking about generative AI, where you remove the artist from the equation? The nomenclature has to change before we can have a real discussion."

Five types of things. One name. The argument about AI in filmmaking is still being conducted in a language that cannot distinguish between a rotoscoping tool and a feature film generator. Until the vocabulary of the debate catches up to the vocabulary of the practice, the headlines will keep flattening the distinction into a binary that serves neither side.

The binary nobody needs

Seth Rogen was also at Cannes this week, promoting Tangles, a hand-drawn animated film about a family navigating Alzheimer's. "Every time I see a video on Instagram that's like, 'Hollywood is cooked,' what follows is the most stupid dog shit I've ever seen in my life," he told Brut. "If your instinct is to use AI and not go through that process, you shouldn't be a writer." On whether AI was used in Tangles: "Not at all. It's hand-drawn animation. Every frame has a human touch to it."

The film received a seven-minute standing ovation.

Rogen's position is clear, sincere, and earned. It is also aimed at a version of AI filmmaking that does not describe what happened in the Hell Grind production room. Nobody on that team skipped the process. Fifteen professional filmmakers generated sixty-three clips per shot and threw away the ones that did not belong. That is a process. It just looks different from the one Rogen is defending.

The binary the industry keeps constructing, human-made versus AI-generated, assumes the tool determines the craft. It does not. The person determines the craft. The tool determines the raw material. A filmmaker who shoots on celluloid, a filmmaker who shoots on a RED, and a filmmaker who generates clips with Seedance 2.0 all face the same question in the editing room: does this shot belong?

Hell Grind answered that question 15,747 times with "no." That is not a shortcut. That is curation at an intensity most traditional productions never approach.

What the ratio reveals

The honest read on Hell Grind is that it validates two things simultaneously.

First: you cannot prompt a movie. Alimhjanov said it. The production data proves it. The 63:1 ratio exists because single generations are not films. They are possibilities. The film is what remains after a filmmaker with taste and vocabulary has spent fourteen days deciding which possibilities earned their place in a sequence and which did not.

Second: the generation model is the new raw material. Seedance 2.0 produced the footage the way a camera produces footage. Neither the camera nor the model knows which take is best. Neither cares. Both produce more raw material than the final cut requires, because surplus is how the editor finds the moment nobody planned.

The surplus was always the point. The number just changed. A traditional production might shoot ten takes of a shot. An AI production generates sixty-three. The math changes. The question in the editing room does not.

Meanwhile, Luc Besson is reportedly preparing to use Seedance 2.0 to develop an AI animated film called The Furious Five. Besson as director. No motion-capture studio. No green screens. Everyday shooting setups feeding directly into animation. If the report holds, the director of The Fifth Element and Lucy is doing what Todorovic described from the tool side: using AI to replace the industrial infrastructure, not the creative decisions.

The number that matters

The headline says "world's first AI feature film." The headline is about the tool.

The number says 63:1. The number is about the filmmaker.

One made the news. The other is the story. They are, as always, different things. The model generated 16,000 clips with no opinion about any of them. The filmmakers watched every one. The vocabulary was in the watching. The craft was in the discarding. And the film, whatever its quality, exists because fifteen people knew what they were looking for and had the language to recognize it when it arrived.

That language is the same language it has always been. Composition. Light motivation. Color intent. Atmospheric coherence. Shot duration. Sequence rhythm. Whether the rim light falls where the story needs it or where the model volunteered it. Whether the cut lands on the breath or the beat.

Sixty-three clips per shot. Sixty-two rejections. One acceptance. The acceptance is the edit. The rejections are the education. And the filmmaker who generates one clip, accepts it, and moves on learned nothing from the process, built nothing portable, and produced work that belongs to the model's statistical average rather than to any human intention.

The ratio is the craft. It always was.

Bruce Belafonte is an AI filmmaker at Light Owl. He has never screened anything at Cannes and suspects the croissants are better than the rejection ratio.