A film entirely created by artificial intelligence was shown in Cannes. $400 thousand went to AI computing power

The 95-minute film «Hell Grind» cost its creators $500 thousand, of which $400 thousand were expenses for artificial intelligence computing power. The whole process took two weeks. The Wall Street Journal provides the details.

Павялічыць

Still from the trailer for the film «Hell Grind»

The topic of artificial intelligence (AI) has been one of the main ones at the Cannes Film Festival for several years now. Previously, discussions around AI were mostly anxious: actors, directors, and screenwriters feared that technology would start replacing people. But this year, as WSJ notes, the mood has become less panicked.

For many in the industry, the question now is only how exactly to use this tool. An example here is the 95-minute film «Hell Grind» (the title can be translated as «Road Through Hell»), created by the San Francisco-based startup Higgsfield AI. In this adventure action film, absolutely everything — characters, locations, lighting, and even small props — is a product of generation.

The film's plot tells the story of four thieves who literally fall into hell. The main character, Rocco, must pass through a dystopian wasteland to save his beloved Lulu.

As WSJ writes, creating such a large-scale visual story took Higgsfield AI only two weeks. The project's budget was $500 thousand, with 80% of this amount going to pay for computing power.

At the same time, the company itself does not develop basic video generation models. Instead, it uses existing systems, such as Google Veo 3, and focuses its efforts on tools that allow images to be consistent and stylistically uniform.

The creators of «Hell Grind» insist that AI does not eliminate the profession of a director. Even with full film generation, a true understanding of cinematography is required. One needs to know the principles of frame composition, shot sequencing, the basics of editing and camera work. AI can draw an image, but it is the human who determines how the scene should look and in what sequence it should proceed.

Lights! Camera! Prompt!

According to Adyl Alimzhanov, Head of Content at Higgsfield, each individual query (prompt) for the model generated about 15 seconds of material. These short clips had to be generated many times, constantly adjusting the query to get the best version. For the first 25 minutes of the film, 16,181 preliminary video generations were needed, from which 253 final shots were ultimately selected.

A particular challenge in creating a feature-length AI film was maintaining visual consistency. AI models can be unpredictable, and in a major film, scenes cannot be allowed to look different in each new shot.

That is why each prompt had to be extremely long and detailed. All parameters were specified: image style, lighting, lens type, camera placement. In addition, the artificial intelligence was reminded of the need to adhere to the laws of physics: «maintaining gravity and inertia — mass must have real weight, correct contact shadows, no props floating in the air».

As a result, on average, each individual prompt consisted of 3 thousand words. Higgsfield even created a separate tool that helps generate such gigantic prompts itself. A user can upload a script page, after which the system will generate a prompt of several thousand words, optimized for creating professional quality video.

Despite the speed of work, the process turned out to be very expensive and labor-intensive. According to the creators, they had to discard hundreds of generated clips due to minor defects: incorrect eye movement of the character, unnatural facial expressions, or an unsuccessful camera movement.

As a result, the company's bill for AI computing power amounted to $400 thousand. To keep costs from rising even further, the creators worked with «non-cloud» providers such as Nebius and CoreWeave, avoiding the services of large technology giants. Such services primarily deal with renting powerful video cards and computing resources for neural networks and often offer more favorable terms for AI projects.

As noted by the production company, their project proved that even in the age of artificial intelligence, filmmaking remains a complex craft. It is impossible to simply ask a neural network to «make a cool 95-minute film» — and get a finished product.

• 3
• 4
• 0
• 4
• 1
• 0
• 0
• 0