Four street thieves are on their way to hell in the action-adventure film, which premiered at the Cannes Film Festival on Thursday. But the compelling thing about “Hell Grind” isn’t the contrived plot: it’s that every character, scene and prop in the 95-minute film was generated by artificial intelligence.
Startup Higgsfield AI took just two weeks to create the film and spent $500,000, 80% of which went toward computational costs.
For the three-year-old San Francisco startup, the finished film is essentially a showcase designed to sell the quality of its artificial intelligence products to Hollywood studios.
Its debut comes as the debate around artificial intelligence and filmmaking is changing — at least at the Cannes Film Festival.
Artificial intelligence has dominated discussions at the annual French Riviera Film Conference for years, with industry insiders and technology experts questioning how much of the filmmaking process – from writing to acting, directing, editing and visual effects – can or should be outsourced to technology, and what that means for jobs and human creativity.
But overall, attendees said the mood this year is shifting from existential dread to cautious acceptance.
At a press conference at the festival last week, actress Demi Moore said actors should find ways to use the technology. “Artificial intelligence is here. So fighting it is a battle we’re going to lose,” she said.
For many in the industry, the question is how to use it.
“My main goal as a filmmaker is I just want to tell stories. In that case, AI can give you the tools to present your story to the world,” said Adilet Abish, Higgsfield’s in-house director and creative producer who worked on Hellraiser.
Adil Alimzhanov, Higgsfield’s head of content, said what might surprise viewers is how much cinematic know-how was required to make the film.
“You have to understand the camera composition, which shots change. Just like you can’t shoot two close-ups in a row, you have to start with an establishing shot,” he said. “You still need those filmmaking skills.”
Higgsfield was valued at $1.3 billion in its latest funding round earlier this year, and its annual revenue run rate topped $400 million in May. It does not make actual video generation models, but relies on existing tools such as Google Veo 3. But it does provide top-level tools to ensure visual effects remain consistent across all next-gens.
lamp! … camera! …quickly!
Alimzhanov said the core of the filmmaking process is prompting the AI model and getting the clips. Each prompt generates approximately 15 seconds of footage. These 15 seconds need to be generated multiple times and the prompts adjusted to get the best version. The first 25 minutes of the movie required the generation of 16,181 initial videos, resulting in 253 final shots.
One of the biggest difficulties in using artificial intelligence to create feature-length films is maintaining consistency in the output. AI models can be unpredictable, and a feature-length movie is unlikely to have scenes that look completely different from one moment to the next.
Therefore, each tip must be very long and detailed. Each one typically begins with a prefix that defines requirements such as style (8k IMAX, photorealistic), lighting (natural light only, “reverse daylight” backlight, shadow side camera), and the type of camera used when shooting (“cine lens,” 180-degree shutter motion blur).
Alimzhanov said lighting is key to avoiding AI sheen, which is often called “slop.” AI-generated videos tend to over-light scenes in unnatural ways.
The prefix must also remind the AI to obey the laws of physics, such as: “Respect gravity and inertia – masses have real weight, correct contact shadows, no floating props.”
Each individual prompt averages 3,000 words.
One aspect that Higgsfield is building and selling to customers is an artificial intelligence tool that can generate these complex, detailed prompts. The user can input pages from the original script and the Higgsfield tool will return a prompt that can be thousands of words long, designed to create production-quality output.
All this prompted the company to raise $400,000 in artificial intelligence computing expenses on the project. However, co-founder and CEO Alex Mashrabov noted that partnering with “neocloud” providers such as Nebius and CoreWeave rather than working with large hyperscale providers will help it control further rising costs.
But constant iteration is a necessary part of the process. “I’ve seen hundreds of videos thrown away because Rocco’s eye twitch wasn’t right, or his jaw wasn’t clenched, or the camera wasn’t moving all the way to the right,” Alimzhanov said. Rocko is the protagonist of Hell Grind as he journeys across a dystopian wasteland to save his fellow thief and love interest Lulu.
“You can’t go into artificial intelligence and say make me a cool 95-minute video.”
Write to Isabelle Busquet: isabelle.bousquette@wsj.com
