Bona Film Group spent two years and tens of millions to produce the first AI theatrical feature film, "Sanxingdui: Future Past"

Written by Geekmovie, author | Qiqi

On the day of the interview last week, Qu Ji Xiaojiang, producer of "Sanxingdui: Future Past," talked about the current state of the production team. Following the theatrical pipeline, they were about to produce the final digital distribution copy and send it out to cinemas nationwide; once the copy is finalized and output, no further changes can be made. Yet the night before the interview, she was still leading the team in redoing shots in the film.

According to her, she "wished she could use the latest model to re-render the whole film the night before release." People on the team often joke about this themselves: other people's live-action films wrap and shut down, but here they are, almost in theaters—why are they still shooting?

This daily routine of overturning and starting over until the very last moment accompanied the entire project for nearly two years.

As the first 100-minute AI hyper-realistic feature film in China to obtain a public screening permit from the China Film Administration and truly reach theaters, "Sanxingdui: Future Past" has no ready-made industry experience to copy.

Why be the first to "eat the crab"? How exactly is a 100-minute AI film made?

Bona Film Group CEO and chief producer of "Sanxingdui: Future Past," Jiang Defu

Before the film officially entered theaters, Geekmovie talked with Jiang Defu, CEO of Bona Film Group and chief producer of "Sanxingdui: Future Past," and Qu Ji Xiaojiang, head of Bona AIGMS Production Center and producer of "Sanxingdui: Future Past," about this two-year pathfinding process.

A 27-year-old film company's "self-revolution"

The starting point of "Sanxingdui: Future Past" is the major historical and cultural subject of "Sanxingdui," which Bona acquired early on.

As a cultural treasure with a 3,000-year history, Sanxingdui has unearthed bronze artifacts right before our eyes, yet left no written records; most of what the world hears are legends.

Facing such a subject, for which it is very hard to determine a creative direction, Jiang Defu admitted that the team initially pondered how to shoot it, because creators cannot simply dominate it themselves—"if the story the creators tell is not what audiences want to hear, that would be quite troublesome."

It happened to coincide with the rise of generative AI technology in 2023-2024, so Bona decided to first make an AI short drama, "Sanxingdui: Future Apocalypse," to test the market and see which direction audiences were actually interested in.

Screenshot from "Sanxingdui: Future Apocalypse"

When the short drama launched on Douyin in 2024, the whole team was in an extremely flexible state of revising while broadcasting. At the time, the platform agreed to launch one episode punctually at 7 p.m. every day, but she was still revising at 6:50, and the platform was so anxious that it called to demand the cloud drive link.

When it reached the fourth episode, they also went to Chengdu to hold a screening for science fiction writers. Because the writers offered many new ideas, the final three episodes were simply overturned and redone entirely.

This AI short drama ultimately garnered over 200 million views, helping Bona gauge audience acceptance of the direction of "modern archaeology and science fiction time travel."

While the short drama was running, Jiang Defu looked around the industry and made an unexpected discovery:

After looking around, we realized we were leading the pack; many traditional film companies have not yet walked the path we have.

Since we have already become the leader, we might as well keep running.

In Jiang Defu's view, the more than 120-year history of film since its birth is itself a history of continuous self-upgrading.

From early film copies that would scratch and become unusable after three or four hundred screenings, to digital copies that can be screened in unlimited cycles and greatly reduce operating costs, every technological iteration has boosted the industry.

So our proactive move today is actually an attitude: opening a brand-new content track and giving audiences one more choice to return to theaters.

However, extending a short drama of a few minutes per episode into a 100-minute theatrical feature is a completely different concept.

A short drama can be revised and launched within ten or twenty minutes, but a theatrical film is a business of selling tickets to audiences; it must pass both content review and technical review, and also leave time for promotion and distribution.

Adding up all this time is terrifying, because different models are iterating competitively; possibly five models are released simultaneously in one month.

While waiting for review and scheduling, watching new models being released constantly, the team could only seize every gap to redo shots.

Now that cloud drives and wireless transmission are widespread in cinemas, the technical conditions for updating at any time exist. "In the future it will become faster and more convenient, and we may also have new ways of controlling content."

Taking two years, from initially having almost no models available to today's technology entering a period of rapid development, Qu Ji Xiaojiang calls "Sanxingdui: Future Past" "a complete record of an era of AI development"—"only this film has had such an experience."

But this work has already begun to produce chain reactions in the industry.

Despite being backed by a short drama with 200 million views, the experience of short dramas cannot be directly copied into a film. A short drama of three minutes per episode emphasizes fast narrative and fast pacing; each episode only needs to leave a small hook. But a theatrical film is a 90-to-100-minute appreciation mode and must have a complete film structure.

Extending a few minutes of excitement to 100 minutes while maintaining consistency in actors' performances, scenes, and overall narrative is a completely different concept.

Many AI short films on the market, to save technical trouble and reduce the AI feel in the image, choose to lay a real person's face over a character. The team abandoned this path from the start, deciding not to rely on any traditional live-action methods at all, not to take any form in real physical space, but to stubbornly pursue native intelligence and directly generate hyper-realistic pure digital human images in virtual space.

Although this is the hardest way, in the long run it can accumulate real digital assets for the company.

Once the archaeologist male lead Jiang Cheng and the astronaut female lead Xiling are established through the story, in the future they can also take audiences to Egypt, the Maya region, the Forbidden City, and other places to continue adventures, gradually accumulating their own life trajectories across one work after another.

To bring virtually generated images into cinemas, they must also pass the hurdle of traditional film companies.

Jiang Defu mentioned that Bona has long maintained a rather particular habit: whether making theatrical films, producing dramas for television stations, or creating short films viewable on mobile phones, all sample footage is invariably screened on the big screen during review.

Because we know very well that no problems can be seen on a phone, but once magnified on the big screen, there will definitely be problems, and all kinds of problems you can't even imagine.

Even so, rapidly iterating models still plunged the team into a cycle of repeated overturning. Often just as one version was finished, a new large model was released, and looking back at previous shots, they felt they weren't good enough, so they had to redo them immediately.

"First, it's not one-click generation; we 'kneaded' it out with a serious attitude."

Jiang Defu did two calculations. First, the production cost: "Sanxingdui: Future Past" involved a team of seventy to eighty people before and after, took more than two years to complete, and had a total investment of over ten million yuan, comparable to the investment of many mid-to-low-budget theatrical films in the tens of millions on the market.

Second is the operating cost of cinemas.

Unlike scrolling short videos on mobile phones with almost zero threshold, when audiences walk into a cinema, the venue rent, projection equipment, staff services, and daily management behind it are all hard expenses, and it still needs to be sustained through box office revenue.

In the past, spending 10 million yuan to shoot a visual effects blockbuster with real people produced rough results, and audiences simply wouldn't buy it; now, using the same 10 million yuan combined with AI technology, from visual effects and macro scenes to the sound environment, everything can be a big notch better than before.

Some filmmakers ask why audiences don't watch. I think don't ask the audience why; we haven't yet done our best. If you want to invite audiences back, we should first do our best, and audiences will come on their own. We need to give audiences more choices.

When the project was just launched two years ago, the first hurdle was that there was no one available.

In the environment at that time, professional filmmakers simply wouldn't use AI to make films, "for them this was a matter of dimensionality reduction, and not just one dimension, but several dimensions." But finding AI technicians wouldn't work either, because they had no film experience at all.

So they could only spend a lot of effort building a cross-disciplinary team, requiring applicants to "at least cross two boundaries first": either fully understand film and love technology, or be pure technical talent but particularly fond of film. The team also brought in quite a few people from game, art, and screenwriting backgrounds.

After two years of running in, because the industry is so short of such experienced hands, Qu Ji Xiaojiang joked that the team has now simply become the industry's "military academy": even if newcomers stay here only a month, they can grasp a composite production process, and many people, after leaving here, quickly became heads or backbone of AI business at other companies.

Besides internal training, the two producers also scoured the entire internet for talent.

On the morning of the visit, as soon as Jiang Defu opened his eyes, he came across an 18-episode AI short drama on his phone and watched four or five episodes in one go.

Actually, its shot production is certainly not up to par; by our standards there are many unprofessional aspects. But this story attracted me. I'm in film; the fact that I could watch four episodes in a row shows its narrative ability is quite strong.

In Jiang Defu's view, Chinese film lacks not only funding but even more screenwriters and directors. Relying solely on the students graduating each year from film academies nationwide is far from enough; some students who got in through the unified exam aren't even very interested in film.

Because large model tools will become simpler and simpler, Bona's core criterion for picking people from short videos is "storytelling ability." After spotting good prospects, the film company then gives them professional experience in feature film production.

Bringing him over to work together is quite labor-saving. He's already very good at short videos and short films; once brought into the team, you just lead him in making feature films and let him know how films are made.

As for whether AI technology will impact existing filmmakers, Qu Ji Xiaojiang feels traditional directors have no need to worry at all, because AI films and live-action films are inherently two different tracks. "The more developed technology becomes, the more people want to see handmade things; this is a kind of human contrarian psychology."

As Jiang Defu said:

Bona Film Group has developed for 27 years and is still young, and must value the future even more. What is the future? Talent plus technology; there must be better screenwriters, better creative subjects, and more composite talent who understand technology, using the ability to master new technology to feed back into our traditional film industry.

In his eyes, from the short drama trial to the completion of the feature film, the biggest gain was figuring out the temper of the new technology. "After going through the entire production process, we know how to use the new technology, and we also know what technical support it can provide for my live-action films and live-action series during use."

On the morning of the visit, Qu Ji Xiaojiang scrolled through the WeChat records on her phone and found a message sent to Jiang Defu on March 9, 2024. That day she sent a test video and told Jiang Defu: "Our character can open his mouth and speak."

The person in the video was none other than Jiang Cheng, the male lead of today's "Sanxingdui: Future Past." Before that, this character was just a static image; his eyes wouldn't move, wouldn't blink, and he couldn't open his mouth, so it could only be advanced with narration paired with text.

Until that day he could finally speak, and although the lip sync and lines still didn't match, popping out word by word, it was precisely this rough test clip that two years later became a 100-minute theatrical film.

Qu Ji Xiaojiang has kept that video to this day: "The excitement I felt then was like seeing a child speak for the first time; that excitement was real."