China's first AI theatrical film "Sanxingdui: Future Past" grosses only one million in its first three days

Three days after its release, the domestic first AI theatrical film "Sanxingdui: Future Past," which netizens had eagerly anticipated, had grossed only one million yuan at the box office.

The film took two years to make, generated more than 1.2 million base images, and involved nearly a hundred people; the production cost disclosed by the filmmakers is in the tens of millions.

"Sanxingdui: Future Past" is not simply a matter of stitching several AI videos together and calling it done. The film runs 100 minutes, and its story spans ancient Shu, ancient Egypt, and the future worlds of 2045 and 2095, and it also had to handle different scenes such as war, ruins, and snow mountains.

From the project's launch in 2024 to its official release, production took two years. Following the model of a traditional film crew, 17 professional roles were set up, including screenwriter, director, art, cinematography, lighting, sound effects, and editing.

Some viewers also approved of the film's visual effects.

The bronze sacred tree, the bronze standing figure, the protruding-eye mask, as well as the ancient Shu royal capital and large-scale sacrificial scenes, all received some praise.

Some viewers also said that after watching, they wanted to visit the Sanxingdui Museum.

Therefore, there is no need to dismiss this attempt as utterly worthless just because of the cold box office.

A character has just discovered a clue and decided to go to the snow mountain; the shot cuts, and the person is already there; after experiencing an avalanche, another shot cuts, and the person is already standing at the entrance to the ruins.

The traveling, searching, waiting, and reactions after encountering danger in between—processes that could have driven the characters' relationships—were all compressed into a few simple shot fragments.

Even more absurd is that in the film, the king of the ancient Shu kingdom 5,000 years ago wants to go to Egypt.

You should know that even today, flying from Sichuan to Egypt still takes 16 hours.

But in the film, after only a few scenes of walking, taking a boat, and riding a camel, the king smoothly meets the Egyptian pharaoh, and the pharaoh even speaks fluent Chinese directly.

(The ancient Shu king and the Egyptian pharaoh in the official promotional video)

This kind of omission looks like lazy editing, but behind it is actually a hard limitation of current video generation models.

By 2026, the length that mainstream models can generate continuously in a single pass still falls in the range of 8 to 30 seconds: Veo 3.1 gives 8 seconds at a time, Kling and most domestic models give 15 seconds, and Seedance 2.5 can reach 30 seconds.

In long-video orchestration work published by Google in September 2026, this kind of problem was given a name: semantic drift and cascading failure, referring to subtle deviations in characters' clothing, props, and lighting in spatial relationships accumulating backward along the shots.

Traveling happens to be one of the hardest types of material to generate: it has no visual spectacle, yet it requires the characters, along with clothing, lighting, and environment, to remain unchanged over a relatively long period of time.

Deleting the process and keeping only the result is the lowest-cost approach at this stage, at the price of losing the space for characters' relationships to grow.

In addition to the disconnects in scenes and settings, the characters' "behavior" and "performance" are also fragmented.

Some film reviews pointed out that the character is still wearing leather shoes and a shirt while hiking in the snow mountain, and after the avalanche is almost completely unharmed, with no change at all in clothing or hairstyle.

(In the official promotional video, the female lead wears a spacesuit almost the entire time)

For humans, an avalanche means impact, burial, cold, and potential injury. After a person experiences these events, their body, clothing, and behavior often change.

But if the generation system does not continuously track these states, different shots may each generate images that look reasonable on their own but contradict one another when put together.

The same goes for characters' emotions: in the previous scene the face is expressionless, but the next shot cuts and the person is suddenly sobbing, with no emotional transition, and the crying merely feels like eyedrops spilling out of the eyes, without even the rims of the eyes turning red.

In addition, in some close-ups, the character's mouth moves, but the other facial muscles lack corresponding changes; these implausibilities make it hard to believe this is a "living person."

The two are not equivalent to film generation technology, but they face a common core problem: the system must not only recognize what is in front of it, but also understand as much as possible the spatial relationships between objects, the laws of motion, and what consequences an action will bring.

03 Why is the film and television industry in such a hurry to make AI films?

First, both technology and industry need an opportunity to prove themselves.

Whether a 100-minute feature can be made means whether AI can enter more complex content production processes.

For film companies, this is also an opportunity to explore cost structures, production efficiency, and new production workflows.

Second, it is capital's and the market's expectation of "the first."

After the release date of "Sanxingdui: Future Past" was announced, Bona Film Group's stock price once rose significantly.

There is also a very realistic consumption issue.

AI-generated short dramas on Douyin exceeded 200 billion views in September 2026, showing that consumers do not inherently reject AI content.

But when there are a large number of AI videos available to watch for free online, why would people still spend 19.9 yuan, 39.9 yuan, or even more to go specifically to a movie theater to watch an AI film?

This article comes from the WeChat public account "Crazy Tech Hall," author: Kong Yilin, published by 36Kr with authorization.