Higgsfield Made a $2M AI Film With Licensed Celebrity Likenesses

The Cully Hill Boys is a 110-minute action-comedy generated entirely with AI, and it just became the most detailed public blueprint for AI filmmaking ever released. The film features AI versions of Israel Adesanya, Quinton "Rampage" Jackson, N3on, and Matt "MK" Kiatipis , and it marks the world's first fully AI-generated feature film to star licensed AI celebrity likenesses. Every prompt, character sheet, location asset, and production log is now public on Higgsfield's platform.

A film school in a project page

It took a 28-person team, including nine directors led by Aitore Zholdaskali and Adilet Abish, to adapt Timothy Planagan's original award-winning screenplay into the film. The startup produced the 110-minute movie in four weeks. The digital cast brings to life the story of three struggling East London rappers whose plan to shoot a music video unravels when a stolen boat full of loot lands them between rival crime syndicates.

The production log alone has 137 entries. Every iteration was tracked: what changed, what version, what verdict. That discipline is part of what Higgsfield is now handing over to the public, alongside the CINEDANCE skill , a structured prompt system built specifically for the Seedance video model on Higgsfield's Canvas.

The economics are hard to ignore

In just four weeks, the group , 40 percent of which had no prior AI experience , reportedly produced nearly 1,000 assets, as well as 100 AI-generated locations. The project cost a total of $2 million. For context, the Forbes report notes that independent films of comparable ambition typically demand one to two years and budgets above $20 million. Higgsfield's own previous feature, Hell Grind, was a 95-minute film produced for about $500,000 requiring roughly 100 assets , meaning the team multiplied production scale roughly tenfold in a matter of months.

About half the $2M budget went to AI compute. The rest covered the human team: directors, editors, costume designers, generative artists, and post-production work that no model handles on its own.

How they actually built it

The entire film runs on text-to-video only , no image-to-video, no starting frames. Every shot is born from reference assets plus text. That constraint is intentional: it's what makes shots cut together consistently. Here's the core production system they developed:

• Character sheets: Three panels per actor , full body front, full body back, close portrait in 3/4 view. The full-body faces are removed so the model only has one place to pull a face from: the close portrait. Two close-ups are made (with and without a smile) to prevent the model from inventing teeth.
• Voice as a locked condition: Register, tempo, accent, and manner are written out precisely and pasted into the audio field verbatim, every single time. Not even a synonym changes. Phonetic accent notation (e.g., th going to f and v, glottal stops) is embedded directly in the line.
• State-specific assets: A character in a clean jacket is a different asset from the same character post-fight with a split brow. Mixing states in one prompt causes the model to blend them randomly across shots.
• Location kits: Locations are generated with anchors , a specific lamp, a specific door , so the model doesn't reinvent the room from different angles. The key trick: generate a slow camera walkthrough of an empty location, and the model draws the other sides of the room to match.
• The 10-15 iteration rule: If a shot hasn't landed after 15 tries, the problem isn't the wording. Split the shot in two, remove an action, or change the angle entirely.

Music was handled by recording the track first, cutting it into 12-second blocks on the vocal's breaths, and passing each block as an audio file labeled as the live performance , not as a guide track. The model won't perform a song if you explain it's a reference; it performs it if you tell it that's what's happening.

The one exception to the text-to-video rule: a fight scene inside a Cadillac that refused to generate cleanly. Two bodies in constant contact, a gun changing hands , limbs fused, the weapon kept disappearing. So the team filmed stunt performers in a real car on an iPhone, then used that footage as motion reference. One day of phone footage replaced a week of failed iterations.

The legal layer is the real innovation

The 110-minute movie aims to be the first AI-generated production featuring legally licensed celebrity likenesses. Each cast member provided high-resolution photographs and voice recordings under agreements covering compensation, permitted scope of use, and script approval before production began. Higgsfield told Forbes that all source assets were permanently deleted within 30 days of wrap and that its agreements bar the company from retaining, training on, or licensing that data to third parties.

The 118-day SAG-AFTRA strike produced Hollywood's first AI consent provisions, requiring advance notice and compensation before a performer's digital replica can be used. Congress is now moving to give that principle the force of law via the NO FAKES Act of 2026 (S. 4591), which would create a federal intellectual property right over every individual's voice and visual likeness. What labor won at the bargaining table and lawmakers are writing into statute, The Cully Hill Boys converts into commercial practice.

The business model implication is significant: the agencies representing N3on and Kiatipis now offer AI likeness licensing alongside sponsorships and endorsements, treating a client's AI avatar as a managed line of business. A performer's likeness went from something to defend to something to license.

What the premiere crowd actually thought

Reactions at the New York premiere were mixed in the way you'd expect from a genuine first. One media professional told Inc. the film "exceeded expectations of what it could eventually be." A video editor in attendance was more precise: some frames came in below 30 FPS, and the AI-generated faces in close-up looked "a little too perfect." He praised the color grading, audio, and graphics, and concluded that for the average viewer, it worked , especially as a first.

The Forbes reviewer had a different experience: the evening featured opening remarks from Higgsfield AI CEO and Co-Founder Alex Mashrabov, followed by a screening and post-premiere reception. That reviewer noted that just a few minutes in, they stopped looking for technical errors and were swept into the story , which is probably the more meaningful benchmark.

What this opens up

The exclusive screening included a technology demonstration of the upcoming Cinema Studio 4.0 and an early look at Higgsfield's AI-powered filmmaking roadmap. The film itself is not headed for a conventional theatrical release , it was built to demonstrate what the platform can do for studios, agencies, and brands.

The open-source release changes the calculus for independent creators. The five rules the Higgsfield team distilled from the production are now public knowledge:

1. Assets first. Nothing generates until every face, place, and prop is locked.
2. Describe everything, every time. The model has no memory.
3. Change one thing at a time.
4. Give the model less freedom , a corner, not a room; a map, not guesswork.
5. Shot won't land? Simplify the shot, not the words.

The assumption that feature-length AI filmmaking requires either a massive studio or a massive budget has been stress-tested and disproved. Around 40 percent of the team had no prior AI experience , they came from traditional production backgrounds and were onboarded through Higgsfield's free public curriculum. That's the part worth sitting with: the craft transferred faster than the industry assumed it would.