Sanxingdui’s AI-made feature meets its first market test

Released across China on October 23, “Sanxingdui: Future Remnants” is reported by several outlets as the country’s first all-AI theatrical feature to receive a public-screening license; it runs 100 minutes. By midday on day four, Maoyan showed ¥1.192M cumulative gross, 8,469 shows that day, 0.1% occupancy and 0.1 viewers per show. Beacon’s midday reading a day earlier was ¥914K — two platforms, two cutoffs, one trend: a ¥125K-plus opening pushed the total past a million yuan, then day four slowed sharply.

Capital markets moved ahead of audiences. After the release date was set, Bona Film Group hit a one-way limit-up on September 28 and two more on October 9 and 12; on the evening of October 12 the company filed a volatility notice flagging “uncertainty over the film’s box-office returns”. On October 26 the stock closed limit-down, off 10.06% at ¥5.54. In the same report, Beacon’s model projected a final gross around ¥2.505M.

• Check Beacon’s terms, opening-day shows and the financials ↗
• See the October 26 Maoyan reading and the share-price report ↗

The timeline itself runs from source to life. Bona set up its AIGMS production center at the end of 2023; after Sora turned the industry in early 2024, it experimented first with “Sanxingdui: Future Revelations”, a 13-episode AI short drama that launched on Douyin that July, passed 200 million plays, and kept being revised while airing — after a Chengdu science-fiction writers’ screening, the last three episodes were rebuilt from scratch. The feature began later in 2024 and took more than two years. ITHome reports 17 departments and nearly a hundred participants, modeled on a traditional crew; Sohu’s making-of account gives a team of seventy to eighty and a total budget in the tens of millions of yuan — comparable to a modest live-action release. The work adds up to more than 1.2 million base images and over a hundred full-film iterations.

Several choices shaped the method. The team refused to graft AI faces onto real actors or shoot in physical locations, generating pure digital humans inside a virtual space instead: the male lead Jiangcheng and female lead Xiling are reusable digital assets, ready for adventures in Egypt, among the Maya, or in the Forbidden City. Every cut was reviewed on a cinema screen — dropped frames, sound-picture drift and subtle goofs hide on small displays. Shots were still being redone days before release; producer Qu Xiaojiang said she “wished to re-render the whole film with the newest model the night before the premiere”. On March 9, 2024, she sent chief producer Jiang Defu the first test clip: “our characters can open their mouths and speak.” Before that, they had only been still images.

• Read the making-of: two years, tens of millions, “you get what you pay for” ↗
• Check the release report and the predecessor short drama ↗

The studio’s division of labor is explicit: “humans lead, AI executes.” Creative direction, narrative logic, Sanxingdui archaeology and humanistic judgment stayed with film-makers and historical advisers, while AI handled generation and execution against the industry’s shared problems — stability and smoothness, facial micro-expressions, physical and temporal continuity. Early on, almost no usable model existed; later they iterated “month by month, sometimes several models at once.” Preparation, set construction and acting direction all happened inside the in-house “Boka cloud studio.” The film carries a dedicated on-screen notice that the picture and some of the sound are AI-generated.

The real bottleneck was people. The hiring bar was to “cross at least two borders” — film-makers resisted AI, technologists didn’t know film — so the team recruited storytellers from games, art, screenwriting and short video, and trained them into “AI tool users with a director’s mind”; peers joked that it had become the industry’s “academy.” Jiang Defu’s arithmetic gives a reference point: ¥10M used to buy a rough live-action effects film, with big set pieces costing millions per shot and requiring overseas work; the same budget plus AI can lift the results considerably. Distribution also cut ticket prices on purpose to bring audiences in with feedback — Bona did not set a high first-film gross as its primary goal.

• See the “humans lead, AI executes” framing at announcement ↗
• Read premiere-night audience feedback and studio quotes ↗

What has clearly been run through: the pipeline, the licensing path and the digital-asset route. Original digital characters sidestep real-actor likeness rights, and the character assets belong to the studio for reuse; more than a dozen countries have reached out about distribution, Malaysia released day-and-date with the mainland, and Russia and France are in talks. Audiences produced genuine positives too — Cover News recorded premiere-night reactions like “the visuals are beyond expectations” and “the two arrow-volley scenes were genuinely stunning.” For teams wanting to tell long AI stories, Qu Xiaojiang calls the film “a complete record of an era of AI development.”

What remains unproven is just as clear. Occupancy of 0.1% says image generation is not the same as a reason to buy a ticket — some in the industry argue AI characters struggle to earn emotional investment; a blunter reading: “if the script fails, the film still flops; this is a market for content.” The financials offer a sober reference: in the first half of 2026 Bona’s AI film-and-drama business earned ¥1.9323M, 0.42% of ¥458M total revenue, at a 96.83% gross margin — generation is indeed cheap (operating cost just ¥61.3K), but the revenue side has not opened up; the company itself wrote that “the technical path, business model and market acceptance still need continued validation.” A second film, “Off-Plan: Guanghan Operation,” is already shooting; ST Huayi’s “Euler’s Identity” and the AI slate inside China Film’s second youth-program phase are on the way. This “first” reads more like a starting line than a conclusion.

• Next: in AI short drama, even audio continuity needs manual checks →
• See AI-video application cases and production notes →