Peter Diamandis opened the discussion with a production budget: 110 minutes of film, 28 people, four weeks and roughly $2 million. About half the money went on compute—the cost of running the AI models that generated the footage.
Those were the figures presented for Higgsfield’s The Cully Hill Boys on Moonshots with Peter Diamandis. A clip followed, featuring an argument about smuggled spices. Higgsfield’s own film page describes an action-comedy about three struggling London rappers who become entangled in a drug war.
Diamandis described the film as using licensed celebrity likenesses. He contrasted its reported budget and schedule with his estimate of $20 million to $100 million and a year to eighteen months for a conventional feature with celebrity talent.
Alex, one of the panelists, had skimmed through the film and watched a substantial portion. It was not his favorite genre; he called it “sort of British Bollywood.” The content interested him less than the execution: “But at the functional level, it is really interesting.”
The discussion turned to how much cheaper that execution could become—and what filmmakers actually receive when a company publishes the materials behind a finished movie.
A production you can study, not freely reuse
Diamandis said Higgsfield had “open sourced all 10 steps in their workflow so anyone can replicate it.” Guest Emad Mostaque described an 80-page guide to how the film was made, calling it “almost open-sourced.”
The attached terms are narrower. Higgsfield’s Originals License 1.0—Study Only covers a production build containing prompts—the instructions given to the models—along with chains of prompts, settings, model selections, workflows, character references, and intermediate and final generations.
That gives filmmakers material to examine beyond the finished images: which models and settings were used, how instructions were linked together, and how the production progressed through generated versions. The license permits viewing, analysis, teaching and discussion, plus limited credited excerpts for genuine editorial or educational purposes.
It does not grant permission to reproduce the workflows, make derivative works, train models on the material, or reuse performers’ likenesses and voices. Nor does it clear third-party rights for viewers. A filmmaker can study how the movie was assembled without acquiring permission to reproduce the production or use its cast.
Which costs could fall next
Diamandis asked what the reported $1 million compute bill might look like by late September. Mostaque’s answer began with a condition: it depends on the quality required and the kinds of shots being made.
The panel said the film used Seedance 2.5. Mostaque described it as a large, expensive model, costing roughly $3 per 30 seconds of generation and accepting dozens of audio and video inputs. A finished film requires many generated shots; its running time alone does not determine the bill.
Cheaper models, he estimated, cost ten to twenty times less while delivering about 90% of the quality. That was his informal comparison, not a measured score. He also contrasted Seedance’s ability to generate 30-second clips with his estimate that a typical shot in current Hollywood releases lasts about three seconds. Model choice and shot design, in his account, leave room to spend less without using the most expensive option everywhere.
Mostaque suggested a comparable film could probably be made for $100,000 in compute, falling to about $10,000 by the new year. Those were forecasts for generation costs, not estimates of the entire production budget: computation accounted for only half of the original reported $2 million.
LTX-2.5 offered another example of where savings might come from. Its documentation describes an open-weight model, meaning its learned parameters are available to download, that generates synchronized video and audio on consumer graphics processors.
Its adaptive rendering allocates computation according to scene complexity rather than treating every scene alike. Its multishot generation aims to keep characters, surroundings, lighting, style and voices consistent across cuts within a single generation. A distilled version—designed to reduce the computation needed—offers another route to cheaper output. These features address both the expense of generating footage and the difficulty of making separate shots belong to the same film.
A face is not just another production asset
The panel’s discussion of performers was less settled. One speaker doubted that the biggest stars would license their likenesses and predicted producers would hire lookalikes instead. Another suggested estates could become eager suppliers of deceased actors’ likenesses, even imagining a guild for dead performers.
These were proposed business strategies, not established legal permissions. Hiring a lookalike does not, on the evidence discussed, establish a right to exploit another performer’s identity. An estate’s willingness to make a deal likewise does not settle which rights it controls.
The conversation also raised a dispute over where a performer’s identity ends and a studio-owned character begins. A panelist reported concerns from film stars about perceived gaps in the SAG-AFTRA agreement. His account conveyed their objections, not a definitive interpretation of the contract. For viewers of Higgsfield’s production materials, the immediate rule is explicit: access does not include permission to reuse performers’ faces or voices.
Generation still needs direction
Diamandis ended with a smaller, more awkward example of the creative problem. He had taken State Street’s executive team through the studio’s “holodeck,” where an AI called Amber offered to help create any movie, song or code. What did they want to make?
The answer, he recalled, was a hip-hop song with no words.
The invitation was “way too open-ended,” Diamandis said. His proposed remedy was to create an environment or movie scene first, draw the visitor into it, and then let them guide what happens next. Rather than asking someone to imagine anything at all, give them something concrete to direct: “People just don’t know how to even start.”